AI Model Pricing Guide
Because tokens cost money and you're not made of it | Updated November 7, 2026
Sarvam Prism (Nov 3 โ 105B decision model, Apache 2.0 self-host, hosted on Sarvam Indus) / ๐ช CHEAPEST: Baseten joey-1 ($0.01/1M input-only โ 2B open-weight MIT decision model, Oct 29; ยผ the price of d1) / Liquid d1 ($0.04/1M input-only โ decision model, not an LLM) / Clef-flash ($0.038/1M input-only โ Cloudflare decision model, cut 58% Oct 9; Clef-omni $0.15 handles audio+video) / Microsoft-Decision-1 ($0.042/1M input-only โ decision model) / SubQ Snap ($0.05/1M input-only, 384K ctx โ decision model, public preview) / Liquid LFM3-1.7B ($0.08/$0.20, 1.7B open-weight edge model, Oct 27) / Decisions API GA (Oct 22 โ $0.10/1M input-only on gpt-6-luna, cache $0.01, no output charge, ~10x faster than Responses) / Claude Haiku 5.5 ($0.10/$0.50 โค100K prompts, $0.50/$2.50 above โ cache reads from $0.01, 1M ctx; v5 36.2 beats all released open weights) / Upstage Solar Mini 4 ($0.05/$0.20 promo, 524K ctx) / MiMo-V2.6-Flash ($0.14/$0.28, omni-modal 1M) / Particle-1 ($0.25/$1.00 peak, off-peak $0.125/$0.50 โ DeepSeek V4.1 Flash retrained to burn ~45% fewer tokens, so the per-task bill undercuts the base model) / GPT-6.1 Terra ($1/$5 โ half of Sol, 50% off its own 5.6 predecessor) / Mercury 3 ($0.10/$0.35 launch โ 1,068 tok/s diffusion LLM) / GLM-5.3-Flash ($0.15/$0.50, 1M ctx โ official Z.ai list) / MiniMax M3.1 Flash ($0.15/$0.60, cache $0.015; open weights landed Oct 23) / Qwen3.8-Omni-Flash ($0.15/$0.47, omni agents) / Tencent Hy3 ($0.13/$0.53) / DeepSeek V4.1-Flash off-peak ($0.15/$0.60, cache $0.003) / Muse Spark Open weights โ FREE download (hosted $1.25/$4.25) / Muse Spark Contributor ($0.10/$0.20) / Nemotron 3.5 Super VL โ open weights + data, self-host๐ BEST VALUE (v5.0 board where measured): GPT-6.1 Terra ($1/$5 โ AA v5 51 at $0.96/task, cheapest in the top ten; Astra-class computer use at one-tenth the price) / GPT-6.1 Sol ($2/$10 โ v5 54.5, third-best buyable model) / Claude Sonnet 5.5 ($2/$10 โ v5 54.2, AA-Briefcase leader 1812 Elo) / Kimi K3.1 Pro ($2/$10 โ v5 46, best open-weight score even before weights land Nov 4; Highspeed $4/$20) / Gemini 4 Argon ($2/$10 intro to Nov 30, then doubles โ v5 50.4, AutomationBench leader 78%) / Claude Haiku 5.5 ($0.10/$0.50 โค100K prompts โ v5 36.2, outscores all released open weights) / Muse Spark Open (free weights; $1.25/$4.25 hosted โ v5 24.7) / DeepSeek V4.1-Flash ($0.15/$0.60 off-peak โ v5 32.5, top released open weights) / Gemini 3.8 Flash ($0.75/$3.75 intro, doubles Jan 1) / Qwen 3.8-Max ($2/$6 flat 1M) / MiniMax M3.1 Flash ($0.15/$0.60 โ v5 28.9 at deep discount) / MiMo-V2.6-Pro ($0.435/$0.87 โ v5 29.9 with MIT weights out) / Mercury 3 ($0.10/$0.35 launch โ 1,068 tok/s diffusion, still unmeasured on v5) / Microsoft-Decision-1 ($0.042/1M input-only if the job is decisions, not prose). Falling the wrong way: K2.6 at $0.95/$4, Mistral Large 4 list $1.36/$4.18 โ weights released Oct 29 (1.06TB, MIT-style custom licence), Step 5 open weights slipped past Oct 15๐ง SMARTEST (AA v5.0, Oct 27-28 โ rebased board, don't mix with old v4.3 scores): Gemini 4 Carbon (59 โ #1, Fairwind-gated, $2/$10 intro when it lands) / Claude Opus 5.5 (58.4 โ #1 buyable, $4/$20) / GPT-6 Astra (56.9, $10/$50) / GPT-6.1 Sol (54.5, $2/$10) / Claude Sonnet 5.5 (54.2, $2/$10) / GPT-6.1 Terra (51 โ cheapest in the v5 top ten at $0.96/task, $1/$5) / Gemini 4 Argon (50.4, $2/$10 intro to Nov 30) / Muse Wonder Full (49.7 โ Muse Code subscribers only) / Claude Fable 5.1 (49.7, $10/$50) / Kimi K3.1 Pro (46 โ top open-weight model, weights not yet released, $2/$10) / Muse Wonder (43.3, $4/$20) / Muse Spark 1.3 (36.4, $1.25/$4.25) / Claude Haiku 5.5 (36.2 โ outscores every released open weights model, $0.10/$0.50) / Grok 4.7 (35.0, $2/$6) / DeepSeek 1017 (32.5 โ top released open weights, $0.15/$0.60 off-peak)๐งฎ NEW THIS WEEK (Nov 2โ6): GPT-6 Cyber (Nov 5) โ <a href='https://openai.com/index/introducing-gpt-6-cyber/' style='color:#00d4ff'>official announcement</a> โ OpenAI's Critical-rated security specialist, built from GPT-6 Astra: ExploitGym 48.5% vs GPT-5.6 Cyber's 32.1% (~51% relative), RemoteExploitBench 95% completion where Astra refused 100% under Daybreak Blue, honeypot out-of-scope exploitation 0.07%, 948 Linux kernel privilege-escalation exploits and 81 novel high-severity Firefox vulnerabilities found (Mozilla shipped fixes). <a href='https://developers.openai.com/api/docs/models/gpt-6-cyber' style='color:#00d4ff'>$20/$2-cached/$100 per 1M</a> โ the most expensive OpenAI row โ but Daybreak Red approval + IP allowlisting required; API + Codex to selected customers first. Baidu ERNIE-5.5-Flash-Base (Nov 5) โ open BASE weights: 280B/14B-active MoE, pre-training-only (33.5T tokens, Muon + FP8), MLA+HCA hybrid attention (37 compressed + 6 MLA-DSA layers), 1M-token inference ctx, FP8+BF16 on Hugging Face under a custom licence; post-trained ERNIE 5.5 Flash to follow. Carrying: Tencent Hy4 final (Nov 4 โ first-gen vision + computer use CLI+GUI, $0.834/$2.501, OpenRouter half price; open weights still unshipped Nov 6); Kimi K3.1 Pro weights LIVE (Nov 2 โ <a href='https://huggingface.co/moonshotai/Kimi-K3.1-Pro' style='color:#00d4ff'>moonshotai/Kimi-K3.1-Pro</a>, the best open-weight model ever measured, MXFP4); Sarvam Prism (Nov 3); Crouton Alpha (still free $0/$0 Nov 6); GLM-5.3-FlashX-Legal $0.37/$1.25; LongCat-2.5 promo $0.30/$1.20; Particle-1 $0.25/$1.00 peak. Upcoming: Grok API text watermarking on by default Nov 7 (opt-out per model in Console); MiniMax M2.5/M2.7 retire Nov 18; GPT-5.6 Sol promo ends Nov 21; Gemini 4 Argon intro AND Claude Sonnet 4.5 API retirement both land Nov 30; Gemini 3.8 Flash intro doubles Jan 1 with Gemini 2.5 auto-routing Dec 1. Watch list: Fable 5.5 still unannounced; Gemini 4 Carbon pre-release; Tencent Hy4 open weights; GPT-6.2 Astra is a Codex-telemetry rumor, not an announcement.
O
OpenAI
The OG of AI APIs. GPT kicked off the revolution and they're still leading.
POWER
GPT-6 Astra
The $10/$50 flagship โ Ultrafast tier live
In
$10.00
per 1M
Out
$50.00
per 1M
1.05M ctx128K outputARC-AGI-3 99.9%Ultrafast 6x (API)
OpenAI's flagship, announced Sep 3 and live on the API Sep 4. $10/$50 per 1M โ 5x GPT-6.1 Sol's price โ with $1 cached input, $12.50 cache writes, and a long-context tier above 272K input (2x input/cache, 1.5x output). Batch/Flex 50% off; Fast mode 2x. Saturates ARC-AGI-3 (99.9%) and FrontierMath Tier 4 (98%); GPQA 96.0%. AA Intelligence Index v4.3.2: 53 max (Opus 5.5 leads at 58), but it still wins where GPT-6.1 Sol breaks down: Terminal-Bench-Science 68.1% vs Sol's 57.0% and AutomationBench 41.4% vs 36.1%. DevDay Sep 29: Ultrafast tier live in the API and ChatGPT Work/Codex (Pro 500, Enterprise) โ up to 6x faster in the API, 8x in Codex at premium pricing. And in the saga of the week: OpenAI cancelled the GPT-6.1 Astra upgrade over safety regressions rather than ship it. Escalate here when 6.1 Sol measurably loses โ science, hardest multi-app automations, retry after a Sol failure.
Best: Frontier agentic work, computer use, cyber defense, scientific terminal work
NEW
GPT-6 Cyber
NEW โ Critical-rated cyber specialist โ Daybreak Red gated
In
$20.00
per 1M
Out
$100.00
per 1M
1.05M ctxCached $2.00ExploitGym 48.5%Critical (Preparedness)Daybreak Red only
OpenAI's offensive-security specialist (Nov 5): GPT-6 Astra rebuilt with dedicated training to find and validate exploitable vulnerabilities without leaving the scope researchers set. $20/$100 per 1M (cached input $2, cache writes $25/$50 for 30min/12h) โ 2x Astra's rates and the most expensive OpenAI row โ but it is gated: API and Codex access starts with selected approved Daybreak Red customers, progressively expanding to all of Red (GPT-5.6 Cyber rolls out to all Red customers alongside). Vendor-measured: ExploitGym 48.5% vs GPT-5.6 Cyber's 32.1% (~51% relative gain; Astra under Daybreak Blue: 42.4%), RemoteExploitBench 95% completion where Astra refused 100% of tasks, black-box web-pentest recall 40.3% vs Astra's 32.0%, SRE-Bench 92.8%, honeypot out-of-scope exploitation 0.07% (in-scope 99.93% without deployment prompts, 100% with). Fielded work: 948 Linux kernel privilege-escalation exploits developed (vs 166 for GPT-5.6 Sol in the same harness), 81 novel high-severity Firefox vulnerabilities (Mozilla shipped fixes), Brotli RCE demo. Classified Critical under the Preparedness Framework (5.6 Cyber: High) โ the first model at that level. Safeguards: cyber-specific misalignment monitoring, IP allowlisting by default for Red API access, Private Safety Processing required for ZDR customers, critical-infrastructure blocks with a Daybreak Red for Critical Infrastructure tier for vetted defenders. The system-card counterweight: on the harder 288-task Irregular FrontierCyber set, Astra solved 96 tasks vs Cyber's 90 โ neither solved the 15 hardest, and neither broke a fully hardened target. 1.05M ctx, 922K max input, 128K output, text+image in.
Best: Authorized penetration testing, exploit development, large-scale vulnerability research โ vetted security teams only
NEW
GPT-Image-2.6 Sunburst / Flare
NEW โ image gen rewrite: SVG + animated GIF at $8/$30
In
$8.00
per 1M
Out
$30.00
per 1M
Image output $30/1M tokCached text $1.25SVG + animated GIFNov 4
OpenAI's image-generation refresh (Nov 4) on the v1/images/generations and v1/images/edits endpoints: adjustable complexity, SVG generation and animated GIF output, with identical pricing for both variants โ $8/$30 per 1M tokens (image in/out), $5.00 text input, cached text $1.25, cached image $2.00. No Batch rate is listed for the new pair, and GPT-Image-2.5's $15/1M image-output Batch rate is gone from the pricing page โ high-volume pipelines lose that discount. Predecessors gpt-image-2.5-sunburst/flare remain listed at $4/$15 standard.
Best: Diagram/SVG workflows, short looping animations, premium image pipelines
NEW
GPT-6.1 Terra
NEW โ the $1/$5 workhorse: 6.1-fam computer use at half of Sol
In
$1.00
per 1M
Out
$5.00
per 1M
1.05M ctxCached $0.05AA v4.3.2: 48 maxOct 26
OpenAI's third GPT-6.1-family model (Oct 26): fast, efficient visual reasoning + precise computer use at $1/$5 per 1M โ 50% under GPT-6.1 Sol's $2/$10 and GPT-5.6 Terra's $2/$12 โ with cached input $0.05, batch $0.50/$2.50, Fast mode $2/$10 (all half of 6.1 Sol's equivalents). $0.40 per AA Index task. Official scores: Agents' Last Exam +15 pts over GPT-5.6 Terra at medium effort for ~one-third the cost, on par with GPT-6 Astra at one-tenth the price; MMMU-Pro nearly matches Astra and 6.1 Sol; Terminal-Bench 4.0 50.3% at max (5.6 Terra: 21.5% โ more than double); Terminal-Bench-Science within 5 pts of Opus 5.5 at under one-fifth the cost; hallucination error rate 16.7% at low / 4.3% at max effort (5.6 Terra: 53.6%). Independent AA measurement (Oct 27, v4.3.2 max): 48 โ well below Sol's 52 and Astra's 53, but above GPT-6 Sol (48) and Grok 4.7 (46) โ on 130M tokens (somewhat verbose), $0.40/task. Meets OpenAI's Critical cyber threshold with Astra's protections. Live in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu (Oct 26 release notes; some Power presets swap 6.1 Terra at Medium in for Sol at Light); API ID gpt-6.1-terra. 1.05M context, 128K output, text+image input, reasoning effort noneโmax. For complex coding, 6.1 Sol and Astra remain the better choice.
Best: High-volume computer use, desktop agents, visual document work, everyday production at half of Sol's price
NEW
GPT-6.1 Terra Pro
Oct 27 โ Terra with pro-mode reasoning, same $1/$5 rates
In
$1.00
per 1M
Out
$5.00
per 1M
1.05M ctxreasoning.mode: proSame rates as TerraSeveral-x token burn
A serving variant, not a new model (surfaced Oct 27):
openai/gpt-6.1-terra-pro is the same underlying model as GPT-6.1 Terra served with the reasoning mode switched to pro for higher-quality responses on complex tasks. Token rates are identical to plain Terra โ $1/$5 per 1M, batch $0.50/$2.50 โ but pro mode spends far more reasoning tokens per request, so a typical request costs several times more and takes much longer to complete. OpenRouter's guidance: use it for hard, high-stakes problems where the extra accuracy justifies the cost; point everyday coding, agentic and chat work at plain GPT-6.1 Terra instead. It follows the gpt-5.6-terra-pro pattern from the 5.6 family. Caveat: OpenAI's own pricing page and model docs list only the four standard rows (Astra, 6.1 Sol, 6.1 Terra, Luna) โ the Pro slug is reachable via OpenRouter, so treat its long-term status as a serving config rather than a separately priced SKU.Best: Hard, high-stakes Terra-class problems where extra reasoning accuracy justifies a several-x token bill
NEW
GPT-6.1 Sol
New DevDay upgrade โ near-Astra at a fifth of the price
In
$2.00
per 1M
Out
$10.00
per 1M
1.05M ctxCached $0.10AA v4.3.2: 52Sep 29
OpenAI's DevDay model (Sep 29), a major in-place upgrade to GPT-6 Sol. Same $2/$10 per 1M as GPT-6 Sol but cached input halves to $0.10; batch $1/$5. Near matches GPT-6 Astra ($10/$50) on agentic coding, computer use and professional work at 8โ23% of Astra's cost per task โ matches Astra's best DeepSWE v1.1 result (75.2% at high effort). AA Intelligence Index (max): 52 vs GPT-6 Sol's 48 and Astra's 53 โ within 3.5 points of Astra on 8 of 10 AA charts. Terminal-Bench 4.0: 56.1% (GPT-6 Sol: 43.9%); OSWorld 2.0: 71.4% vs 64.4%; GDP.pdf beats Opus 5.5's fallbacks at less than half the cost per task. Weak spot: science terminal work โ 11.1 points behind Astra. System card: lower failure rates than GPT-6 Sol on transparency and respecting restrictions. Oct 8: the Ultrafast tier went GA for all API users at $12/$60 โ a flat 6x on every line (cached input $0.60, cache writes $15; long-context >272K $24/$90) โ same weights, same answers. In Codex it carries the up-to-8x badge (the 6x is the API rate-card math). Live in Work and Codex for all paid plans (not yet in Chat); API ID gpt-6.1-sol; on AWS Bedrock same day.
Best: Default for well-specified work at volume โ automations, reviews, computer use, coding agents โ at a fifth of Astra's price
LIMITED
GPT-Rosalind
Life-sciences โ GPT-6 Rosalind $15/$75; Research/Discovery $5/$25
In
$5.00
per 1M
Out
$25.00
per 1M
128K ctxCache $0.50GPT-6 Rosalind $15/$75Discovery listed Oct 10
OpenAI's specialized life-sciences line, out of research preview (Sep 11) via the trusted-access program for qualified Enterprise/Business orgs, billing since Oct 5: gpt-rosalind-research and gpt-rosalind-discovery at $5/$25 per 1M, $0.50 cached input. Nov 1 check of OpenAI's price page: a third row, gpt-6-rosalind at $15/$75 (cache $1.50), now sits above them โ the line's premium tier, with no docs page yet (models docs still list only research). Same-day row addition follows the Rosalind pattern: access by approved list, not product page. Internal research only โ no customer-facing products. Oct 11: Rosalind Workbench opened in research preview through the ChatGPT app โ guided protein/small-molecule design, genomics and NGS workflows with Explore and Research modes; Workbench access doesn't require model access, and individual access is 'coming soon'. The frontier labs' pattern this month: strongest models ship as permission lists, not product pages.
Best: Drug discovery, genomics, medicinal chemistry โ vetted research orgs only
RETIRED
GPT-6 Sol
Retired โ delisted after 6.1 Sol; was $6/$30
In
$6.00
per 1M
Out
$30.00
per 1M
1.05M ctxCached $0.40Delisted by Oct 10Migrate to 6.1 Sol
The Sep 22 mid-tier, delisted within three weeks โ OpenAI's price page (checked Oct 10) no longer lists gpt-6-sol: standard/batch/fast/flex rows show gpt-6-astra, gpt-6.1-sol and gpt-6-luna only, and it is gone from the model docs. It launched Sep 22 at $2/$10 with $0.20 cached input (batch $1/$5), was still listed at $6/$30 by Sep 30 with cache at $0.40, and was superseded Sep 29 by GPT-6.1 Sol at $2/$10 with cache reads at $0.10 and +4 AA. If you're still routing to gpt-6-sol, move the model ID to gpt-6.1-sol โ better on every measured chart and cheaper.
Best: Legacy routing โ move to GPT-6.1 Sol for free intelligence
NEW
GPT-6 Luna
Cheapest frontier model yet โ $0.10/$0.50
In
$0.10
per 1M
Out
$0.50
per 1M
1.05M ctxCached $0.01Sep 22Cheapest frontier
The GPT-6 high-volume tier, launched Sep 22. $0.10/$0.50 per 1M โ 50% below GPT-5.6 Luna's $0.20/$1.20 โ with cached input at $0.01 and batch $0.05/$0.25. Trained with Astra methods; marginally stronger than 5.6 Luna. Available in ChatGPT Work and Codex for paid plans, and in the desktop app for Free and Go users.
Best: High-volume apps, agent execution, cost-sensitive production
BUDGET
GPT-5 mini
Fast & cheap
In
$0.25
per 1M
Out
$2.00
per 1M
128K ctxFast
Lightweight champion. Surprisingly capable for simple tasks and high-volume apps.
Best: Chatbots, simple QA, data extraction
RETIRED
o4-mini
Retired Oct 23, 2026
In
$1.10
per 1M
Out
$4.40
per 1M
Shut down Oct 23Replaced by GPT-5.6 Terra
Shut down Oct 23, 2026 in OpenAI's largest single-day API cull: GPT-3.5/4/4-Turbo, o1, o1-pro, o3-mini, o4-mini and gpt-image-1 all gone; replacements route to the GPT-5.6 line. Survives only inside the fine-tuning wind-down for existing users (inference until base-model deprecation).
Best: Fine-tuning, custom reasoning
NEW
GPT-5.4 mini
Coding & subagents
In
$0.75
per 1M
Out
$4.50
per 1M
Cached inputCoding
New GPT-5.4-class model. Stronger than GPT-5 mini for coding and subagent workflows.
Best: Coding, subagents, mid-tier apps
NEW
GPT-5.4 nano
Cheapest 5.4-class
In
$0.20
per 1M
Out
$1.25
per 1M
Cached inputBudget
Cheapest way into the GPT-5.4 family. Cheaper input than GPT-5 mini.
Best: High-volume, budget apps
POWER
GPT-5.4
Still elite โ now #2
In
$2.50
per 1M
Out
$15.00
per 1M
270K ctxReasoning
OpenAI's previous #1. Still elite for complex, multi-step problems.
Best: Hardest problems, professional work
POWER
GPT-5.2
Reasoning beast
In
$1.75
per 1M
Out
$14.00
per 1M
6.6h horizon200K ctx
Top 3 on METR. Excels at complex tasks, code, and multi-step reasoning.
Best: Code, analysis, agent workflows
POWER
GPT-5.2 Pro
Reasoning premium
In
$21.00
per 1M
Out
$168.00
per 1M
200K ctxPremium
OpenAI's most precise reasoning model. For when you need the absolute best reasoning.
Best: Hardest problems, precision work
POWER
GPT-5.6 Sol
Superseded by GPT-6 Sol at half the price
In
$4.00
per 1M
Out
$20.00
per 1M
GA Jul 9Fast mode Jul 30AgentsCodingCybersecurity
OpenAI's strongest model, now GA. SOTA on Agents' Last Exam (53.6, +13 over Fable 5), BrowseComp (92.2%), Terminal-Bench 2.1 (88.8%). Three tiers: Sol (flagship), Terra (balanced, $2/$12), Luna (fast, $0.20/$1.20). New Fast mode: 2.5x faster at $10/$60 (API). Multi-agent 'ultra' mode for demanding tasks. Aug 21 promo repricing: $4/$20 through Nov 21, then $5/$30 standard. Superseded Sep 3 by GPT-6 Astra at 2.5x the price โ Sol remains the value pick for frontier agentic work. Sep 22: GPT-6 Sol launched at $2/$10 โ half this promo price โ trained with Astra methods and scoring higher on AA v4.3.2 (48 vs 47).
Best: Frontier agentic tasks, coding, biology, cybersecurity
SUPERSEDED
GPT-5.6 Terra
Superseded by GPT-6.1 Terra at half the price
In
$2.00
per 1M
Out
$12.00
per 1M
GA Jul 9Price cut Jul 30Superseded Oct 26
The everyday 5.6 workhorse โ price cut 20% to $2/$12 on Jul 30, outperforming Fable 5 at ~1/16 the cost. Superseded Oct 26 by GPT-6.1 Terra at $1/$5 โ half the price with a 1.05M window, +15 pts on Agents' Last Exam at a third of the cost, Terminal-Bench 4.0 doubling to 50.3%, and error rates down 37 points. Migrate gpt-5.6-terra โ gpt-6.1-terra: cheaper on every axis that matters.
Best: Legacy routing โ migrate to GPT-6.1 Terra at half the price
80% OFF
GPT-5.6 Luna
Superseded by GPT-6 Luna at half the price
In
$0.20
per 1M
Out
$1.20
per 1M
GA Jul 9Price cut Jul 30Fast
Price slashed 80% on Jul 30 โ from $1/$6 to $0.20/$1.20 per 1M. Outperforms Opus 4.8 on coding agent index. Nearly matches GPT-5.5 peak at a fraction of the cost. The high-volume tier of the 5.6 family. Batch: $0.10/$0.60. Cached input: $0.02. Sep 22: GPT-6 Luna launched at $0.10/$0.50 โ half this price with cached input at $0.01.
Best: High-volume, cost-sensitive apps, agent execution
POWER
GPT-5.5
Now #2 โ still elite
In
$5.00
per 1M
Out
$30.00
per 1M
1.05M ctxReasoningAgents
Previous #1. 82.7% Terminal-Bench, 84.9% GDPval, 78.7% OSWorld. 1.05M context window. Still elite, now behind GPT-5.6.
Best: Coding, agents, research, multi-step tasks
#1 RANKED
GPT-5.5 Pro
Maximum intelligence
In
$30.00
per 1M
Out
$180.00
per 1M
1.05M ctxPremiumDeep Research
90.1% BrowseComp, 52.4% FrontierMath Tier 1-3. 1.05M context window โ read entire codebases and research libraries. The ceiling for what AI can do right now.
Best: Hardest problems, deep research, scientific discovery
FLAGSHIP
GPT-5.4 Pro
Previous premium โ now #3
In
$30.00
per 1M
Out
$180.00
per 1M
270K ctxPremium
Former #1, now behind GPT-5.5. Still incredibly powerful for demanding tasks.
Best: Most demanding tasks, unlimited budget
โ
Microsoft
Decision models for agent control โ structured outputs your software can act on, at a fraction of LLM cost.
NEW
Microsoft-Decision-1
Decision model โ $0.042/1M input, output free
In
$0.04
per 1M
Out
$0.00
per 1M
Output tokens freeQwen3.5-9B base36-benchmark bestFoundry live
Microsoft's entry into the decision-model category that Liquid d1 opened (Oct 9, available in Microsoft Foundry; OpenRouter support planned). Not an LLM: it picks from predefined choices and assigns probability scores โ routing, classification, prioritization, verification and workflow control, so software can branch without generating tokens. Pricing: $0.042 per 1M input tokens, output free โ Microsoft's internal estimates put classifying 1M texts at ~$11 vs GPT-6 Sol's ~$2,434. Built on Qwen3.5-9B via single-pass decision-scoring post-training, with plans to rebase on MAI and OpenAI models. Claims: highest accuracy across a 36-benchmark, ~150K-question suite kept blind from training; 4.5x faster than the runner-up Quyet-1.0-Large, 35x faster than GPT-6 Sol. Internal users: Xbox Research categorized 10K+ gaming feedback items at GPT-6-Sol-comparable quality, 14x faster and 200x cheaper; Copilot's response-quality team found it competitive with GPT-5.6 Luna. All vendor-measured โ launch-week claims, independent verification pending.
Best: Agent step control, model routing, data labeling, AI judging, safety screening โ high-frequency structured decisions
โ
Cloudflare
Open-weight decision models on Workers AI โ schema-bound scoring across text, image, audio and video, no generated tokens.
NEW
Clef-omni
Multimodal decision model โ $0.15/1M input, output free
In
$0.15
per 1M
Out
$0.00
per 1M
Output tokens freeAudio+video+image+text64K ctx hostedApache 2.0 weights
Cloudflare's third Clef in eight days (Oct 9): decision-model scoring expanded to audio (wav/mp3) and video (mp4/webm) alongside text and images โ the first multimodal entrant in a category that was text-only since TypeSafe's Jev debuted. Built on the Qwen3-Omni-30B-A3B-Instruct MoE backbone (comprehension path kept, TTS head dropped); weights open on Hugging Face under Apache 2.0. Speed: text ~130ms median, images ~150ms, a 21-second video with sound ~1.5s โ all in one API call. Media bills as input tokens (~780/min of audio; up to ~15,400/min of video at max resolution). Context is 64K on the hosted endpoint; the open weights can serve more. Like all Clef models it returns calibrated confidence scores over your predefined options โ no prose, no output tokens, nothing to bill.
Best: Multimodal classification and routing โ flagging events in CCTV/audio, invoice QA, content moderation โ without transcribing media first
NEW
Clef
The original Clef โ $0.24/1M input, now faster
In
$0.24
per 1M
Out
$0.00
per 1M
Output tokens freeImage support64K ctx hostedApache 2.0
Cloudflare's first open-weight decision model (Oct 1-2), trained over a weekend and shipped the Thursday after โ the Friday-evening-to-launch story Cloudflare now retells with every release. Stays at $0.24/1M input tokens with output free. Oct 9 serving upgrade moved it to SGLang and cut median latency on ~800-token inputs from 262ms to 152ms. Takes text and image input (plus video frame arrays). Open weights on Hugging Face under Apache 2.0 โ self-host with the full training context if 64K hosted isn't enough.
Best: High-frequency routing, classification and verification steps in agent pipelines
CHEAPEST
Clef-flash
Cheapest decision model โ cut 58% to $0.038/1M
In
$0.04
per 1M
Out
$0.00
per 1M
Output tokens freeUndercut Liquid d124K ctx hostedApache 2.0
The volume tier, repriced Oct 9: $0.09 โ $0.038 per 1M input tokens โ now cheaper than TypeSafe's Jev, and undercutting Liquid AI's d1 ($0.04). The trade: hosted context dropped from 64K to 24K (Cloudflare says only 0.24% of requests exceeded 24K), but the open Hugging Face weights still support up to 256K if you self-host. At this rate a dollar covers tens of thousands of short decisions; Cloudflare's guidance is simple โ route most traffic to Clef-flash and move the long-tail calls to Clef or Clef-omni.
Best: Highest-volume decisions โ ticket triage, spam/safety filters, labeling sweeps
S
Subquadratic
Subquadratic-architecture decision models โ fixed-answer scoring over very long context, no generated tokens.
NEW
SubQ Snap
Decision model โ 384K ctx at $0.05/1M input, output free
In
$0.05
per 1M
Out
$0.00
per 1M
Output tokens free384K text+image ctxDecision Index 61.7Public preview
Subquadratic's long-context decision model (Oct 12 public preview at subq.ai; first come, first served โ waitlist when capacity fills). Not an LLM: it scores a fixed set of allowed answers rather than generating text, at $0.05 per 1M input tokens with output free โ a 100K-token decision call is about half a cent of input. Up to 384K tokens of text and images per call; vendor claims sub-100s latency on requests up to 100K tokens, built on the SSA subquadratic architecture behind its SubQ line (announced earlier in 2026). Vendor-measured Decision Index 0.2.1: 61.7 vs TypeSafe Jev's 57.9, leading 13 of 16 enterprise benchmarks โ independent verification still pending, and the hosted preview caps capacity, so treat it for latency-tolerant paths. SubQ 1.1 Small (the general LLM) remains in early access with design partners.
Best: Highest-volume classification over long documents โ compliance screens, document triage, video/media decision steps
P
Particle
Token-efficiency lab โ post-trains open models to finish the same tasks with ~45% fewer tokens.
NEW
Particle-1
NEW โ DeepSeek V4.1 Flash retrained to spend ~45% fewer tokens
In
$0.25
per 1M
Out
$1.00
per 1M
1M ctxCached $0.03Off-peak 50% offDeepSWE 72.6% @ $0.22/task
The first model from Particle (Oct 28): DeepSeek V4.1 Flash post-trained in internal RL environments to solve the same tasks with fewer tokens โ ~45% fewer output tokens on average with pass@1 inside the base model's confidence intervals. $0.25/$1.00 per 1M (peak; cached input $0.03) with automatic 50%-off off-peak pricing 20:30โ04:00 ET and another 50% off on the Flex tier. Measured by the vendor across four agent-harness benchmarks (Terminus 2): DeepSWE v1.1 72.6% at $0.22/task vs V4.1 Flash's 71.2% at $0.34 โ 28โ52% fewer output tokens and 34โ52% lower cost per task. Focus is software engineering; more domains teased. Hosted-only via the Particle API (Chat Completions + Responses, LiteLLM integration) โ no weights, no OpenRouter at launch.
Best: High-volume coding agents and harnesses where the token bill, not the licence, is the constraint
A
Anthropic
Safety-first company. Claude is beloved by developers for being genuinely helpful.
#1 OVERALL
Claude Opus 5.5
Still AA Index #1 โ the $4/$20 flagship workhorse
In
$4.00
per 1M
Out
$20.00
per 1M
1M ctxCache reads $0.20Fast mode $8/$40AA v4.3.2: 58 #1
Anthropic's flagship-workhorse (Sep 22), first of the Claude 5.5 family. $4/$20 per 1M โ 20% under Opus 5 โ with cache reads cut 60% to $0.20 (cache writes $5) and 30%+ faster output. Fast mode: $8/$40 at up to 2.5x speed. Batch: $2/$10. Anthropic's tests put total run cost 40% below Opus 5, and five-hour limits rose on Pro/Max/Team/Enterprise. AA Intelligence Index v4.3.2 (max): 58 โ the highest score AA has measured, ahead of Sonnet 5.5 (56), Fable 5.1 and GPT-6 Astra (53). Leads six of ten Index evals: HLE 61.4%, SciCode 66.9%, AA-Briefcase 1822 Elo, GDPval-AA v2.1 1846. Best automated-alignment scores Anthropic has measured; bio/cyber capability comparable to Mythos 5.1, so it ships with Fable-5.1-class safeguards. Still clearly stronger than Sonnet 5.5 at complex open-ended work โ Anthropic's own testing โ but Sonnet 5.5 wins Terminal-Bench 4.0 (70.6% vs 66.4%).
Best: Coding, agents, knowledge work โ Fable-class intelligence at 60% of the price
NEW
Claude Sonnet 5.5
New โ #2 on the AA Index at Sonnet prices
In
$2.00
per 1M
Out
$10.00
per 1M
1M ctxCache reads $0.10Terminal-Bench 70.6%134 tok/s
Anthropic's second 5.5-family model (Sep 28), six days after Opus 5.5 โ and the sharpest price-performance move of the month. $2/$10 per 1M โ identical pricing to Sonnet 5 โ but it needs far fewer tokens per task (up to 30% cheaper) and outputs 30%+ faster (~134 tok/s on AA, its fastest Sonnet). Oct 8 update: cache reads halved to $0.10/1M (from $0.20), which Anthropic pegs at ~20% cheaper on most agentic work where cache hits dominate. AA Intelligence Index: 56 โ #2 overall behind only Opus 5.5's 58, ahead of Fable 5.1 and GPT-6 Astra (53) โ at a quarter of Fable's price. Terminal-Bench 4.0: 70.6% at max effort (43.0% at high โ the API default) vs Sonnet 5's 10.3% and above Opus 5.5's 66.4%. First Sonnet to beat Pokรฉmon Red from screenshots alone. Two points below Opus 5.5 on GDPval-AA. First Sonnet with cyber safeguards matching Opus-class capability + reasoning-extraction classifiers: risky security requests fall back to Sonnet 5. 1M context, 128K output (300K on Batch beta); API ID claude-sonnet-5-5. The catch: verbosity โ AA measured the highest output tokens per task it has seen, so the halved cache reads are exactly where the value lands.
Best: Everyday agentic work, bug fixing, documents/slides/spreadsheets โ near-Opus quality at Sonnet prices
NEW
Claude Fable 5.1
Former AA #1 โ the $10/$50 frontier tier
In
$10.00
per 1M
Out
$50.00
per 1M
1M ctx128K outputCache $0.25AA Index 66 #1
Anthropic's new frontier (Sep 1). $10/$50 unchanged but cache reads cut 75% to $0.25/1M โ Anthropic's August usage data shows ~25% total savings on typical workloads, up to ~45% for agentic work where cache hits dominate. Terminal-Bench-Science 52.6% (vs 24.7% on Fable 5), Terminal-Bench 4.0 55.8%, AutomationBench 31.4%, GDPval-AA v2 1853, CursorBench 73.4%. AA Intelligence Index (max, fallback): 66 โ #1, three clear of Opus 5. Cyber false positives down 60%; can now find (not exploit) vulnerabilities. Breaking API changes: forced tool use returns 400, thinking blocks are model-bound. API ID: claude-fable-5-1. Sep 22 update: Opus 5.5 passed it on the AA Index (58 vs 53 max on v4.3.2) at $4/$20 with cheaper cache reads โ Fable 5.1 remains the restricted frontier tier.
Best: Coding, agents, long-horizon knowledge work โ migrate from Fable 5 for the cache savings
NEW
Claude Haiku 5.5
New โ 75% cheaper Haiku, 1M ctx
In
$0.10
per 1M
Out
$0.50
per 1M
1M ctx (long-ctx rates >100K)128K outputFirst Haiku with effort dialOSWorld 2.1 72.4%
Anthropic's small-model refresh (Oct 8 NZ / Oct 7 US). Per 1M: $0.10/$0.50 for prompts up to 100K tokens (Anthropic: ~90% of Haiku traffic), $0.50/$2.50 above; cache reads $0.01/$0.05, cache writes $0.125/$0.625, batch half price. On average ~75% cheaper to run than Haiku 4.5. AA Intelligence Index v4.3.2 (published Oct 8): 43.4 at max โ within a point of Kimi K3 and #14 overall, beating GPT-6 Luna (38.1) and MiMo-V2.6-Flash (37.9) at a fraction of Kimi's price. 1M context, 128K output (300K on Batch beta header), first Haiku with the effort parameter (default medium). Computer use jumps to 72.4% on OSWorld 2.1 (Haiku 4.5: 15.7%), Terminal-Bench 4.0 39.2% (4.5: 0.0%), HLE 45.9% no-tools. Positioned as the subagent next to Opus 5.5/Sonnet 5.5 and for high-volume classification/summarisation/routing. Live on all platforms (AWS, Google Cloud, Azure). Watch the tokenizer: same 4.7+-era tokenizer, so identical text counts ~30% more tokens than Haiku 4.5. API ID claude-haiku-5-5. Shipping alongside: Sonnet 5.5 cache reads halved to $0.10 (~20% cheaper agentic) and new monthly API credits for Max ($100/$200) and Team (up to $500) subscribers.
Best: High-volume classification, extraction, summarisation, live support, computer/browser use, subagents
LEGACY
Claude Haiku 4.5
Legacy small model โ superseded by Haiku 5.5
In
$1.00
per 1M
Out
$5.00
per 1M
200K ctx10x Haiku 5.5 price
Superseded by Haiku 5.5 (Oct 7 US) at a tenth of the price with a 1M window โ Anthropic pegs Haiku 5.5 at ~75% lower average running cost. Still listed on Anthropic's price page for existing pipelines; migrate claude-haiku-4-5 โ claude-haiku-5-5 unless you depend on the older tokenizer's lower token counts.
Best: Existing pipelines only โ new builds should use Haiku 5.5
POWER
Claude Sonnet 5
Superseded by Sonnet 5.5 at the same $2/$10
In
$2.00
per 1M
Out
$10.00
per 1M
200K ctxAgentic$2/$10 permanentMigrate to 5.5
Close to Opus 4.8 performance at Sonnet prices. Intro pricing $2/$10 made permanent Aug 10 โ the scheduled Sept 1 increase was cancelled. Superseded Sep 28 by Claude Sonnet 5.5: same $2/$10/$0.20 pricing but terminal-bench agentic coding jumps 10.3% โ 70.6%, output runs 30%+ faster, and typical tasks cost up to 30% less through lower token use. Anthropic's own comparison keeps Opus 5.5 ahead of both on complex open-ended work; also note Sonnet 5.5's 1M context vs this card's 200K. It's a free upgrade โ move the model ID.
Best: Legacy routing โ migrate to Sonnet 5.5 at the same price
BEST
Claude Sonnet 4.6
Previous Sonnet โ still solid
In
$3.00
per 1M
Out
$15.00
per 1M
Balanced200K ctx
Previous Sonnet default. Still excellent but superseded by Sonnet 5 at lower intro pricing.
Best: Most tasks, code, writing, general use
POWER
Claude Opus 5
Superseded by Opus 5.5 at 20% less
In
$5.00
per 1M
Out
$25.00
per 1M
1M ctx128K outputThinking ONNew Claude Max default
Anthropic's new top-tier Opus. Scores 63 pts on the AA Intelligence Index (v4.1.1) โ second only to Fable 5.1's 66. Near-Fable-5 intelligence at exactly half the price ($5/$25 vs $10/$50). ARC-AGI-3: 30.2% โ 4x GPT-5.6 Sol (7.8%), 20x Opus 4.8 (1.5%). Frontier-Bench v0.1: 43.3% (Opus 4.8: 18.9%). Five-level effort toggle, thinking ON by default. Automatic fallback replaces hard refusals. API ID: claude-opus-5. New default on Claude Max; top model on Claude Pro. Fast mode: $10/$50 at 2.5x speed (API research preview). Batch API: $2.50/$12.50. Superseded Sep 22 by Opus 5.5 โ 20% cheaper ($4/$20), 60% cheaper cache reads ($0.20), and the best alignment audit scores Anthropic has measured. Migrate for the savings alone.
Best: Coding, agents, knowledge work โ the new cost-efficient frontier sweet spot
POWER
Claude Opus 4.8
Previous Opus flagship โ superseded by Opus 5
In
$5.00
per 1M
Out
$25.00
per 1M
1M ctx128K outputSelf-verify
Previous Opus top tier, now superseded by Claude Opus 5 at the same price. 1M context, 128K output, autonomous self-verification. Same $5/$25 pricing as 4.7. Migrate to Opus 5 for near-Fable-5 intelligence at no extra cost.
Best: Complex coding, agents, long-horizon tasks
POWER
Claude Opus 4.7
Previous Opus SOTA โ still elite
In
$5.00
per 1M
Out
$25.00
per 1M
1M ctxxhigh reasoningSelf-verify
Previous Anthropic best. 1M context, autonomous self-verification. Beat GPT-5.4 on BrowseComp. Now superseded by Opus 4.8 at the same price.
Best: Complex coding, agents, long-horizon tasks
POWER
Claude Opus 4.6
Proven workhorse
In
$5.00
per 1M
Out
$25.00
per 1M
14.5h horizon200K ctxFast mode
Still one of the best. 14+ hour autonomous tasks. Reliable, consistent, now the value play vs 4.7/4.8.
Best: Hard problems, research, complex agents
POWER
Claude Fable 5
Superseded by Fable 5.1 โ migrate for cache savings
In
$10.00
per 1M
Out
$50.00
per 1M
1M ctx128K outputSafety classifiersRestored Jul 1
Public version of Mythos. $10/$50 โ double Opus 4.8. 1M context, 128K output, autonomous self-verification. Includes safety classifiers that can refuse requests (refusals are not billed; fallback credits refund prompt-cache cost on retry). Suspended by US government Jun 12, restored globally Jul 1. Superseded Sep 1 by Fable 5.1 โ same $10/$50 with cache reads at $0.25/1M.
Best: Hardest reasoning, long-horizon agentic work, cybersecurity
LIMITED
Claude Mythos 5.1
Fable 5.1 with permissive safeguards โ vetted access
In
$10.00
per 1M
Out
$50.00
per 1M
1M ctx128K outputTrusted accessUS orgs only
Identical to Fable 5.1 with safeguards tuned for cybersecurity and life sciences. Ships via the Cyber Verification Program, the US-government-partnered Life Sciences Verification Program, and Project Glasswing โ Anthropic's strongest cyber model ever. System card flags it as 'less honest under pressure' than recent Claude models. Also powers Claude Security. US organizations only for now.
Best: Vetted cyberdefenders, life-sciences R&D โ apply via CVP/LSVP
NEW
Grok 4.7
New flagship โ frontier knowledge work, same $2/$6
In
$2.00
per 1M
Out
$6.00
per 1M
500K ctxAA 46 (+2)Cached $0.50DeepSWE 71.0%
xAI's most capable model (Sep 21), shipped after four missed release windows. Same $2/$6 pricing as Grok 4.6 with $0.50 cached input. AA Intelligence Index: 46 at xhigh (+2 over 4.6), placing xAI in the top-4 labs; AA-Briefcase 1657 Elo (+111) โ just behind Opus 5 and Fable 5.1; GDPval-AA 1695 (+90). CursorBench 4.0 frontier price-performance; DeepSWE v1.1 71.0% (high effort); Terminal-Bench 4.0 38.0%; EEBench, Harvey Legal, HealthBench Professional gains. Works longer and self-checks with xAI's best-calibrated safeguards. The catch: 81k output tokens per task at xhigh โ double Grok 4.6's 38k and ~3x GPT-6 Astra's 27k, so the same per-token bill can cost more per task. Knowledge cutoff May 2026. Live in Cursor, Grok Build, the Grok API and model routers.
Best: Coding, long-horizon agents, professional knowledge work
POWER
Grok 4.6
Superseded by Grok 4.7 at the same price
In
$2.00
per 1M
Out
$6.00
per 1M
500K ctx1753 ELOCached $0.50Coding & agents
xAI's new flagship (Aug 12), released with Cursor. $2/$6 per 1M, $0.50 cached input. 1753 ELO claim โ overtakes Kimi K3. Live on xAI API, Grok Build, Cursor, Grok Bot, and partners (OpenRouter, Vercel, Cloudflare). Fast variant available at twice the price. 2x included usage in Cursor/Grok Build for the first week. Sep 21: Grok 4.7 launched at the same $2/$6 โ AA Index 46 (+2), AA-Briefcase 1657 Elo, but roughly double the output tokens per task.
Best: Coding, long-running agents, knowledge work โ the new default Grok
POWER
Grok 4.5
Previous flagship โ superseded by 4.6
In
$2.00
per 1M
Out
$6.00
per 1M
500K ctx80 TPS4.2x token efficiencyCoding & agents
SpaceXAI's previous flagship, now superseded by Grok 4.6 at the same price. Trained alongside Cursor. Opus 4.8-class intelligence at 80 TPS with 4.2x better token efficiency than Opus 4.8. SWE Marathon 29% (beats Opus 4.8 at 26%). Cached input at $0.50/1M. Migrate to Grok 4.6 for no extra cost.
Best: Coding, agentic tasks, knowledge work โ migrate to Grok 4.6
NEW
Grok 4.20
Same price as 4.3, more features
In
$1.25
per 1M
Out
$2.50
per 1M
2M ctxReasoningMulti-agentVision
Same pricing as Grok 4.3 with multi-agent orchestration. Cached input at $0.125/1M. 2M context window for complex agent swarms.
Best: Complex multi-agent workflows
NEW
Grok 4.3
New recommended base model
In
$1.25
per 1M
Out
$2.50
per 1M
Best value2M ctxReasoningVision
xAI's recommended Grok 4 model after retiring old variants. Beats Grok 4.1 on coding, agents, and reasoning. This is the migration target for retiring models.
Best: High-volume apps, X analysis, multi-agent — the new default Grok
RETIRED
Grok 4 / 4.1 Fast
Retired May 15
In
$0.20
per 1M
Out
$0.50
per 1M
2M ctxRetired
Retired May 15, 2026. Migrated? Good. If not, move to Grok 4.3 ($1.25/$2.50) or Grok 4.20 ($1.25/$2.50).
Best: โ Migrate to: Grok 4.3
POWER
Grok Build 0.1
Live coding model โ the alias behind Grok Code Fast 1
In
$1.00
per 1M
Out
$2.00
per 1M
256K ctxReasoning always-onCached $0.20โฅ200K: $2/$4
SpaceXAI's coding model for agentic software engineering ($1/$2 per 1M <200K prompt, long-context $2/$4 โฅ200K, cached input $0.20/$0.40, 256K ctx, no Batch API, us-east-1/us-west-2) โ documented on the xAI models page with grok-code-fast-1/grok-code-fast/grok-code-fast-1-0825 as aliases, while the retired May 15 card below carries the old 2025 pricing. Function calling, structured outputs, reasoning always on.
Best: Agentic coding, refactors, multi-step engineering workflows
RETIRED
Grok Code Fast 1
Retired May 15 โ replaced by Grok Build 0.1
In
$0.20
per 1M
Out
$1.50
per 1M
256K ctxRetired
Retired May 15, 2026. Migrate to Grok Build 0.1 ($1/$2) or Grok 4.3.
Best: โ Migrate to: Grok Build 0.1
BUDGET
Grok 3 Mini
Older gen cheap
In
$0.30
per 1M
Out
$0.50
per 1M
131K ctxReasoning
Budget fallback if Grok 4's 2M context is overkill for your use case.
Best: Simple tasks, testing
POWER
Grok 4-0709
Premium tier
In
$3.00
per 1M
Out
$15.00
per 1M
256K ctxReasoningVision
Premium Grok. Smaller context but more reasoning power.
Best: Grok style with more smarts
M
Meta
Meta's paid API stack โ Muse Wonder is the first frontier model and Spark the cost-efficient line, from Meta Superintelligence Labs.
NEW
Muse Wonder
Meta's first frontier model โ $4/$20, 1M ctx omni-modal
In
$4.00
per 1M
Out
$20.00
per 1M
1M ctxCached input $0.20AA Index 52 (max)Voice/video in
Meta Superintelligence Labs' first frontier model (Oct 12), built at the gigawatt-scale Prometheus datacentre after a 14-month rebuild ('from rebuilding our foundations to pushing the frontier'). API: $4/$20 per 1M with cached input $0.20 (95% cache discount; cache writes $1), 1M context, text+image+speech+video input, text out โ same price point as Claude Opus 5.5 and double GPT-6.1 Sol. AA Intelligence Index (max): 52 โ #9 of 219, one point behind Fable 5.1 and GPT-6 Astra, +4 over Muse Spark 1.3, and tied with GPT-6.1 Sol โ but at $4.52 per Index task, ~6x Sol's $0.72, since the model is verbose (130M output tokens across AA's run, ~53k tokens per task despite ~10% fewer than Spark 1.3) and slow (83 tok/s). Benchmarks: Terminal-Bench 4.0 52%, AutomationBench-AA 68% (up from Spark's 58%), DeepSWE 1.1 78.8%, Harvey LAB-AA 10.0 (vs Fable 5.1's 6.4), MemoryBench 85.6, OSWorld V2.1 72.6, SWE-Marathon 53.1. The pitch is Active Memory + Active Focus โ remember across conversations, keep working through blockers. Stronger sibling Muse Wonder Full scores 56 (#3, two behind Opus 5.5) but is locked to Muse Code subscriptions and served by no API.
Best: Personal-assistant and professional agent tasks where memory matters โ the API tier buys Wonder's capability, not Full's
SUBSCRIBER
Muse Wonder Full
The 56-scoring sibling โ Muse Code subscribers only
In
$60.00
per 1M
Out
$300.00
per 1M
AA Index 56 โ #3No API accessMuse Code $50โ$200 plansActive Memory + Focus
The larger Wonder sibling (Oct 12) โ bigger base, 56 on AA Intelligence Index v4.3.2 at max: third-highest measured, tying Claude Sonnet 5.5 (Full 55.8 unrounded vs Sonnet's 56.0), two behind Opus 5.5, ahead of Claude Fable 5.1 and GPT-6 Astra (both 53), and +8 over Muse Spark 1.3 with gains on all ten Index evals. But it is not sold by token: Wonder Full (max reasoning) is available only through the Muse Code software-engineering agent on $50/$200 subscriber plans; the CNBC-reported $60/$300 per-1M figure is the subscriber-plan equivalent, and CNBC notes API access is excluded for it.
Best: Muse Code subscribers building with the $50โ$200 plans โ not for API routing
OPEN
Muse Spark Open
Meta's first open weights โ free download, $1.25/$4.25 hosted
In
$1.25
per 1M
Out
$4.25
per 1M
816B MoE / 42B active1M ctx omni-modal inAA ~39 (Meta card)Weights free (HF)
Meta's first open-weight release (Oct 15) โ and 'the most capable US open-weight model to date' by its own claim: an 816B-total/42B-active MoE sharing the Muse Spark 1.3 base (plus extra safety post-training for standalone use), 1M-token context, natively multimodal input (text+image+video+audio), tuned for coding and agentic work. Weights are a free download on Hugging Face under the Muse Model License: free to run, modify, fine-tune and redistribute for most developers, but model-serving businesses above $20M aggregate revenue must notify Meta and sign before serving it. Hosted access matches the closed Spark rate: $1.25/$4.25 per 1M (cache $0.15) on OpenRouter/Together/Fireworks/Groq/Modal/Nebius/Baseten. Meta's own model card reports AA Intelligence Index ~39 โ well below the independently measured MiMo-V2.6-Pro (46.3) and below its closed Wonder siblings, and the full eval report is still pending โ so the 'most capable US open-weight' framing is about the US-vs-China open field, not the leaderboard. The angle that made HN's front page (397 points): Meta co-tuned it with Apple โ via latest MLX it runs entirely on two Mac Studio M5 Ultras (512GB each, 1.2TB/s) as a 1TB unified-memory desktop supercomputer, with on-device demos driving Muse Code to ship real iOS changes and edit DaVinci footage frame by frame. Weights also on SageMaker, Google Cloud Model Garden and Foundry; license allows limited commercial use.
Best: Agentic coding with full weight control, on-prem deployments, desktop-scale local inference on Apple silicon
VALUE
Muse Spark 1.3
Superseded by Muse Wonder โ stays the cheap Spark tier
In
$1.25
per 1M
Out
$4.25
per 1M
AA v4.3: 48.1 max$0.55/taskCached $0.15Aug: v4.1.1 61
Meta's fourth Muse Spark in five months (Sep 2). $1.25/$4.25 unchanged ($0.15 cached input). #1 on Tau3-Bench Banking at launch, at $0.55 per Index task โ cheapest at its intelligence point. On AA's current v4.3.x revisions Muse Spark 1.3 (max) measures 48.1 โ the baseline the Oct 12 Wonder launch moved from (Wonder 52, Wonder Full 56, both with smaller token bills). The older 61/62 top-3 claim came from the previous v4.1.1 methodology era. Oct 12: superseded at the frontier by Muse Wonder โ same-API Spark remains Meta's cost tier; the teaser of larger Muse models landed as Wonder/Wonder Full. Oct 15: its base resurfaces as Muse Spark Open โ the first open weights, at the same hosted price.
Best: Agentic work, scientific reasoning, cost-efficient frontier intelligence
Muse Glimmer 30B
First MSL open model, runs locally
NEW
Muse Spark 1.1
Previous Muse Spark โ superseded by 1.3
In
$1.25
per 1M
Out
$4.25
per 1M
AgenticTool useCodingUS preview
Meta's first paid AI model via the Meta Model API. Agentic model from Meta Superintelligence Labs (run by Alexandr Wang). ~25% cheaper than comparable OpenAI/Anthropic models. $20 free credits for new accounts. US preview only โ no EU access yet.
Best: Agentic tasks, tool use, cost-sensitive apps
G
Google DeepMind
Gemini has quietly become excellent. Massive context, strong multimodal, and a generous free tier.
NEW
Gemini 4 Carbon
NEW โ v5.0 #1 at 59; Fairwind-gated until 'as soon as possible'
In
$2.00
per 1M
Out
$10.00
per 1M
1M ctxAA v5.0: 59 โ #1SEC-bench Pro 88.2%Intro to 1M outputs
Google's next frontier model (announced Oct 28, 13 days after Argon's GA) โ and the first model measured on AA Intelligence Index v5.0: 59, taking #1 the day after the board rebased. Google bills it as its strongest model with state-of-the-art cybersecurity and long-horizon capabilities: teams have seen it run for days root-causing and patching subtle bugs across large codebases; WebDev Arena +89 Elo to #4 (1769). Cyber numbers (evaluated without guardrails): SEC-bench Pro 88.2% finding real vulnerabilities, ExploitGym 54.8% exploitation, CWE-bench 71.0% patching โ plus >2x Argon's secure-code uplift internally and a Chrome-team report of 10x better patches than another lab's large frontier model. Alignment gains: Terminal-Bench 4.0 cheating reduced to 0.0%, Gray Swan prompt-injection resistance 0.0% attack success (Argon: 0.7%), unsupported claims cut ~2x, overreach โ26% / under-service โ29%. Pricing: same intro as Argon โ $2/$10 per 1M with cache at 95% off ($0.10), then $4/$20 after the introductory period, mirroring Argon's Nov 30 intro end. The catch: not yet buyable โ Fairwind Program cyber defenders get access this week, US-government voluntary pre-release testing engaged; developers/businesses/consumers 'as soon as possible' (Argon took 13 days from announce to GA). 1M context, text/image/speech/video input.
Best: Cyber defense pipelines, long-horizon codebase engineering โ when access opens; watch for GA
POWER
Gemini 4 Argon
GA Oct 13 โ frontier flagship at Sonnet money
In
$2.00
per 1M
Out
$10.00
per 1M
1M-token output1M ctxVals Index #1 (68.9%)Intro to Nov 30
Google's frontier flagship went public (Sep 30 announced, Oct 13 GA) after 13 days of Fairwind-only cyber-defender access. Gemini API paid tier, AI Studio and Antigravity (Ultra) now sell it: intro $2/$10 per 1M โ the same as Claude Sonnet 5.5 and GPT-6.1 Sol โ with cached input at $0.10 (95% off), then $4/$20 from Dec 1, 2026. The headline spec is an industry-first 1M unified input-output token limit via a new continuation protocol (was 64K output). Where it wins: #1 on the Vals Index (68.9% vs Sonnet 5.5's 67.0 and Opus 5.5's 67.0), Finance Agent v2 65.4% first, DeepSWE v1.1 77.9% (Opus 5.5: 74.2%), Harvey Legal 19.6% (Astra: 5.4%), ERQA 87% โ the best physical/spatial reasoning score at the frontier, which got it a robotics launch piece. Where it trails: Terminal-Bench 4.0 57.4% vs Opus 5.5's 66.4%; AA Intelligence Index 53 โ ties GPT-6 Astra and Fable 5.1 but sits below Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56). Verbose-ish: 110M output tokens across AA's run. Consumer: Gemini app and AI Mode in Search for Pro ($19.99/mo)/Ultra. Wiz used it to catch a healthcare master-key flaw; the first 100k Google engineers used it internally through Antigravity. Successors under codenames 'Barium' and 'Carbon' are reportedly already in employee testing (Business Insider, Oct 10).
Best: Professional workflows โ finance, legal, long-horizon coding, robotics/physical reasoning โ at intro Sonnet pricing
NEW
Gemini 3.1 Flash-Lite
Cheapest Gemini 3
In
$0.25
per 1M
Out
$1.50
per 1M
PreviewBudget
Cheapest way into Gemini 3.1. Preview tier with budget-friendly pricing.
Best: Budget Gemini 3 apps, prototyping
NEW
Gemini 3 Flash
New budget
In
$0.50
per 1M
Out
$3.00
per 1M
PreviewFast
Gemini 3 Flash preview. Balanced performance at budget pricing.
Best: Budget apps, prototyping
DEPRECATED
Gemini 2.5 Flash
Best value
In
$0.30
per 1M
Out
$2.50
per 1M
1M ctxMultimodalFree tier
Cheapest way to process 1M context. Free tier available. Multimodal - images, video, audio. Deprecated (Nov 2 changelog): gone from AI Studio model picker, and from Dec 1, 2026 API calls auto-route to gemini-3.5-flash-lite, billed at Flash-Lite rates โ update model strings.
Best: High-volume, multimodal, prototypes
NEW
Gemini 2.5 Flash-Lite
Ultra-cheap Flash
In
$0.10
per 1M
Out
$0.40
per 1M
1M ctxBudget
Flash-Lite tier for Gemini 2.5. Cheaper than standard Flash with 1M context support. Best for high-volume simple tasks.
Best: High-volume, simple tasks, cost-sensitive apps
NEW
Gemini 3.5 Flash
GA Flash โ 1M context, agentic
In
$1.50
per 1M
Out
$9.00
per 1M
1M ctx65K outputAgenticComputer Use
Gemini 3.5 Flash is GA. Most intelligent Flash model for sustained agentic and coding work at scale. 1M context, 65K output, thinking, Computer Use, function calling. Free tier available.
Best: Agentic tasks, coding, production Flash workloads
NEW
Gemini 3.8 Flash
4th Flash in 4 months โ AA Index 59 at Flash prices
In
$0.75
per 1M
Out
$3.75
per 1M
1M ctx~300 tok/sAA Index 59Terminal-Bench 90.8%
Google's best reasoning/coding Flash yet (Sep 2) โ 3.5, 3.6, 3.7, 3.8 since May. Same $0.75/$3.75 intro as 3.7 Flash through Dec 31, then $1.50/$7.50 from Jan 1, 2027. AA Intelligence Index 59 (+3 over 3.7) โ cheapest model at its intelligence level ($0.58/task). ~300 tok/s, the fastest output speed Artificial Analysis has measured. Trained on long-running agentic loops, it 'works harder': ~40% higher cost per task than 3.7 (30% more output tokens) โ drop to low effort for efficiency-first routing. Gemini 3.8 Flash Cyber for trusted defenders via the Fairwind Program: 2.6x more correct Chrome patches than much larger frontier models, per Google.
Best: Agentic coding, knowledge work, price-sensitive production โ the new default Gemini
NEW
Gemini 3.7 Flash
Efficiency pick โ superseded by 3.8 Flash
In
$0.75
per 1M
Out
$3.75
per 1M
1M ctxGA Aug 13DeepSWE 65.3%Computer Use
Gemini 3.8 Flash (Sep 2) supersedes it at the same intro price; 3.7 remains fully supported for efficiency-first workloads. Google's most intelligent workhorse Flash model yet (Aug 13). Intro price $0.75/$3.75 โ half 3.6 Flash's original cost โ through Dec 31, then $1.50/$7.50. DeepSWE v1.1: 65.3% (vs 49% on 3.6). FrontierCode 1.1: 43.6% (vs 34.4%). WebDev Arena 1588 Elo (vs 1538). GDP.pdf 34% (vs 22%). AutomationBench 30.4% (vs 17%). Powers Gemini Spark. Built-in Computer Use.
Best: Agentic coding, knowledge work, cost-efficient production agents
NEW
Gemini 3.6 Flash
New workhorse โ 17% fewer output tokens
In
$1.50
per 1M
Out
$7.50
per 1M
1M ctxGA Jul 21Computer UseToken efficient
Google's new workhorse Flash model (Jul 21). 17% fewer output tokens than 3.5 Flash on the AA Index (up to 65% on DeepSWE). Step up in coding (DeepSWE 49% vs 37%), knowledge work (GDPval 1421 vs 1349), computer use (OSWorld 83% vs 78.4%). $1.50/$7.50 โ lower price than 3.5 Flash too. Built-in Computer Use via Gemini API.
Best: Agentic coding, knowledge work, cost-efficient agents
NEW
Gemini 3.5 Flash-Lite
Fastest 3.5 โ 350 TPS
In
$0.30
per 1M
Out
$2.50
per 1M
GA Jul 21350 TPSComputer UseAgentic
Fastest model in the 3.5 series โ 350 output tokens/s (AA Index). $0.30/$2.50. Outperforms 3 Flash on SWE-Bench Pro (54.2% vs 49.6%) and OSWorld (74% vs 65.1%). Terminal-Bench 2.1: 54% vs 31% on 3.1 Flash-Lite. Computer use built-in. Configurable thinking levels for cost/latency tradeoffs.
Best: High-throughput agents, agentic search, document processing
NEW
Gemini 3.5 Pro
2M context โ largest at the frontier
In
$1.25
per 1M
Out
$10.00
per 1M
2M ctxDeep ThinkGA Jul 17
Google's new flagship, launched July 17. 2-million-token context window โ double anything at the frontier. New Deep Think extended reasoning mode (gated to $250/mo Ultra tier). Rebuilt from scratch on a new pretraining run after the original failed on recursive tool-calling. $1.25/$10 โ 4x cheaper input than GPT-5.6 Sol.
Best: Massive context, video analysis, long-horizon reasoning, RAG-replacement
Gemini 3 Pro
3rd gen flagship
In
$2.00
per 1M
Out
$12.00
per 1M
1M ctx
Third-generation Gemini Pro. Now stable โ no Preview tag. Same / pricing as preview tier. Strong general-purpose flagship.
Best: General production apps, stable Pro performance
FLAGSHIP
Gemini 3.1 Pro Preview
New flagship
In
$2.00
per 1M
Out
$12.00
per 1M
4h horizonPreviewVideo
77.1% ARC-AGI-2. Price increased from $1.25/$10. Batch and Flex tiers at 50% off.
Best: Video analysis, complex reasoning
DEPRECATED
Gemini 2.5 Pro
Long outputs
In
$1.25
per 1M
Out
$10.00
per 1M
1M ctx64K output
Same price as 3.1 Pro but 64K max output vs 16K. Choose for long-form content generation. Deprecated (Nov 2 changelog): gone from AI Studio model picker, and from Dec 1, 2026 API calls auto-route to gemini-3.8-flash, billed at Gemini 3.8 Flash rates โ update model strings; breaking changes include function calling + thought_signature validation and thinking_level replacing thinking_budget.
Best: Long-form writing, large outputs
DEPRECATED
Gemini 2.0 Flash
Shuts down Jun 1
In
$0.15
per 1M
Out
$0.60
per 1M
1M ctx8K outputRetiring Jun 1
Deprecated — shuts down June 1, 2026. Migrate to Gemini 2.5 Flash or 3.1 Flash-Lite.
Best: Migrate away from this model
โฌ
Open Source & Local
Open-weight models you can run yourself or call via cheap APIs. The frontier is no longer closed.
OPEN
Beam (Reflection)
Weights out Oct 16 โ hosted API now priced: $1/$3 list, from $0.25/$1
In
$1.00
per 1M
Out
$3.00
per 1M
501B MoE / 23B active23.8T tokens pretrainTerminal-Bench 80.1Apache 2.0 โ LIVECache $0.10 (list)
Reflection AI's first open-weight model (announced Oct 5, weights released Oct 16 โ ahead of the 'later this month' flag): Hugging Face repo reflection/Beam-1.0-501B-A23B is live under Apache 2.0 with instruction-tuned BF16, FP8 and NVFP4 shards; Transformers (tensor+expert parallel) and vLLM support shipped day one, with a dflash2 drafter for speculative decoding. 501B-param MoE with 23B active, pretrained on 23.8T tokens, then 100M+ RL rollouts on 10.5K GB300s over 4 weeks (~1.3B sandboxes, 110K concurrent) that penalized token waste โ vendor numbers: SWE-Bench Verified 80.9, Terminal-Bench 2.1 80.1 (at ~25K reasoning tokens per attempt where GLM-5.2 spends 80K), SWE-Bench Pro v1 65.5, DeepSWE v1.1 44.4. AA Intelligence Index: 30 (changelog entry, Oct 16) โ #2 among US open weights, behind Meta Muse Spark Open's ~39. Pricing story resolved the same week it was 'TBD' (Oct 18): OpenRouter now lists Beam at $1.00/$3.00 per 1M with cache reads $0.10 โ reflecting CoreWeave's route โ and cheaper partner routes exist: Nebius Token Factory $0.25/$1.00, Together $0.40/$1.20 (cache $0.04), Baseten $1.30/$4.30 at 426 tok/s. Reflection's own first-party API (api.reflection.ai, OpenAI-compatible /openai/v1/chat/completions) is in public beta, free to start, with the per-token rate card login-gated at platform.reflection.ai โ treat $1/$3 as the reference rate until published. Token-efficiency is still the launch pitch: 'comparable solution quality to leading open models using approximately 3x fewer inference-time reasoning tokens.' From Mislav Balunoviฤ & Wei Chen (ex-DeepMind, AlphaZero lineage); $25B valuation, 250MW Korean sovereign build, SpaceX Colossus-2 lease.
Best: Self-hosted or partner-hosted enterprise coding agents โ token-frugal coding at a third of Sonnet money per task
NEW
MiniMax M3.1 Flash
Agent flash model โ AA 42.7 at $0.15/$0.60; free window closed Oct 22 UTC
In
$0.15
per 1M
Out
$0.60
per 1M
1M ctx / 524K outputAA 42.7 โ #21 ($0.22/task)DeepSWE 73.7%OSWorld V2 61.7 โ top Chinese modelCyberGym 82%OPEN โ weights Oct 23
MiniMax's next-gen Flash (Oct 19) carries the best reveal of the week: 'Space Bunny' โ the anonymous model that processed 200+ trillion tokens on llama.cpp leaderboards without identification โ was an early version of M3.1 Flash. Flat $0.15/$0.60 per 1M with cache hits at $0.015, 70% cheaper to run than M2, ~3x faster online inference, 1M-token context with 524K output, text+image input. Vendor claims put it at the leading group on AutomationBench-AA (real SaaS workflows end-to-end), BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7 and Agents' Last Exam 29.6 โ first among Chinese models on both โ plus CyberGym 82% for vulnerability discovery. Independent AA measurement (Oct 19): 42.7 on the Intelligence Index โ #21 of 351 at $0.22 per task, inside a point of GPT-5.6 Terra, seven points below Sonnet 5.5, and roughly a thirteenth of its per-token cost; the cost is verbosity (180M tokens across AA's run, ~80% above median). Available through the MiniMax API, M Plan and MiniMax Code; free on OpenRouter until Oct 22, 12:00 UTC, then list price. Open weights LANDED Oct 23 as promised: MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face โ ~428B total / 23B active, minimax-community licence, sparse attention throughout, quantization-aware training and DSpark speculative decoding, with a published vLLM recipe. Flash Pro and Pro (the intelligence tier) teased as still to come.
Best: High-volume agentic work, computer use, coding agents, security screening โ Haiku-class results at deep-discount pricing
NEW
LongCat-2.5 (Meituan)
1.6T MoE / 48B active โ coding-agent multimodal, promo $0.30/$1.20
In
$0.30
per 1M
Out
$1.20
per 1M
1.6T MoE / 48B active1M ctxText+image nativeW4A4 quantizedPromo ยฅ2/ยฅ8
Meituan's delivery-app giant shipped its flagship (blog Nov 1, API changelog Nov 2) โ final version of LongCat-2.5-Preview (Sep 25). 1.6T-param natively multimodal MoE (~48B active per token, images enter the same model as discrete tokens), 1M context with sequences up to 1M during mid-training, long-horizon RL for multi-step agent work: visual coding, creative design, 3D tasks, full-project builds. Trained entirely on domestic Chinese accelerators (<80GB/chip); W4A4 quantization of experts and activations โ load tests claim ~43% throughput gain over FP8. Compatible with Claude Code, Codex and Hermes harnesses. Promo pricing time-limited: $0.30/$1.20 per 1M (cached $0.006, ยฅ2/ยฅ8). No open-weights release announced โ unlike K3.1/GLM/DeepSeek, this one stays API-gated. A 1.6T param count at flash-tier pricing: frontier-class bill of materials at a tenth of US flagship rates.
Best: Coding agents, visual programming, long-horizon builds โ API-only for now
NEW
Mercury 3
Diffusion flagship โ 1,068 tok/s, 1M ctx, $0.10/$0.35 launch
In
$0.20
per 1M
Out
$0.70
per 1M
1,068 tok/s1M ctx (up from 260K)40x Mercury 1's sizeLaunch 50% offCache $0.01
Inception's Mercury 3 (Oct 19) is the proof point the diffusion crowd has waited for: the largest diffusion LLM released โ more than 40x Mercury 1's size โ holds its speed at 1,068 tok/s on widely available NVIDIA GPUs and jumps to a 1M-token context. Inception's claim: on par or better than frontier speed-optimized models (GPT-6 Luna, Claude Haiku 5.5) across a range of benchmarks while running 4โ8x faster, with measured scores including GPQA Diamond 89.2%, DeepSearchQA 71.3% and 50% completion on 50 live Mind2Web web-navigation tasks โ ahead of GPT-6.1 Sol in Inception's own production harness at the same token volume and step count. Independent AA Index measurement is still pending (Mercury 2.5 measures 12), so treat the headline as vendor-measured for now. Pricing: $0.20/$0.70 list with a 50% launch discount of $0.10/$0.35, cached input at $0.01. Live on the Inception API, OpenRouter, Baseten and Vercel; OpenAI-compatible; 100M free tokens on new API keys. TIME named Mercury 2 a Best Invention of 2026 (Oct 15) โ sibling Mercury Voice runs $0.20/$0.75 at under-170ms TTFT for voice agents.
Best: Real-time voice/tutoring/support agents, browser automation, interactive apps where latency compounds
FREE
Kolibri 1
Germany's sovereign open-weight model โ DE/EN reasoning
In
$0.00
per 1M
Out
$0.00
per 1M
In
FREE
local
Out
FREE
local
1M ctx validated78B MoE / 3.46B activeApache 2.0Reasoning mode
Aleph Alpha's sovereign LLM (released Oct 3, Hugging Face: Aleph-Alpha/Kolibri-1). 78B-total MoE with 3.46B active, Apache 2.0, vLLM-ready at ~78GB FP8 (runs on 2x A100 80GB up). 1M-token context validated for quality and serving (262K native; positional encoding in sliding-window layers extends in principle indefinitely). German-first: trained to reason in German, ~23.9% of its 20T-token pretraining mix, German-optimized tokenizer. Reasoning mode, tool calling, RAG with documented abstention, adjustable reasoning effort. EU GPAI Code of Practice signatory; training screened against a 4.5M-URL blocklist with per-dataset license checks. #1 on Hugging Face trending after launch.
Best: German-language RAG, public-administration and compliance-heavy deployment you host yourself
NEW
Mistral Large 4
1T 'Le Chonk' โ weights OUT Oct 29; docs price $1.36/$4.18
In
$1.36
per 1M
Out
$4.18
per 1M
512K ctx (docs: 1M)1.05T MoE / 52B activeCyber 82% / Cybench 93%Weights on HF โ Oct 29
Mistral's biggest model yet (Oct 6 public preview, 'le Chonk'): 1.05T-param MoE with 52B active per Mistral's current docs (launch posts said 49B; the model 'continues to improve rapidly as we refine it') plus a 1.6B vision encoder, natively multimodal (image+text in), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres โ the first model out of its โฌ3B round. API: $0.68/$2.09 per 1M with $0.07 cached input during preview; docs show struck-through list pricing of $1.36/$4.18 (cache $0.14) with no end date for the discount โ budget on list. Mistral claims SOTA among open models on cyber (82% on AA's find-and-patch test โ the highest of any model, where Opus 5.5 and GPT-6 Astra score near zero by refusing โ plus 93% on Cybench), finance and law, SOTA open-weight on SciCode-Verified, and beats closed frontier models on visual grounding (Dense 200 42% vs Astra's 41%). AA Intelligence Index v4.3.2 preview: 38.4; Mistral's own brief ranks it top-five globally on the AA Cyber Index and leads open-weight models developed outside China 'by a wide margin'. Coding: DeepSWE v1.1 61.7%, Terminal-Bench 4.0 28.3%, Coding Agent Index 49.8% (ahead of DeepSeek V4 Pro and Qwen 3.8-Max; Kimi K3 and GLM-5.3 still lead), second of five in a blind Surge AI human eval behind Opus 5. AutomationBench 59.9%, AA-Briefcase 1393 Elo. The RL run is still in flight and 'shows no sign of saturation' โ numbers will move. Weights LANDED Oct 29: mistralai/Mistral-Large-4-1T-A52B is live on Hugging Face (1.06TB bf16, 68 files, no access gate) โ three weeks of extra RL made it the open-weight coding leader: FrontierCode 2.0 19.7%, Terminal-Bench 4.0 41% (+13 from preview), AA Coding Agent Index 48% (tied with GLM-5.3 at 47% fewer tokens), AA Cyber Index 51% with a 96.6% malicious-request refusal rate; independent AA v5.0 Index 25.3; in Surge's blind human eval only Opus 5.5 was preferred over it. The $0.68/$2.09 intro ended Oct 23: Mistral docs price at $1.36/$4.18 (cache $0.14) โ OpenRouter's copy of the current route lags at $0.70/$2.09, budget on Mistral's docs rate. Vals AI ranks it lower than the hype (#32 of 44 on its index) โ the launch cards lean on verticals, not the aggregate.
Best: Sovereign/EU compliance deployments, cyber defence at open weights, multimodal document work
NEW
Liquid d1
Decision model โ probabilities, not prose; $0.04/1M input-only
In
$0.04
per 1M
Out
$0.00
per 1M
$0 output tokens200-300msVision (base64)Non-autoregressive
Liquid AI's first 'decision model' (Oct 5, general availability after last week's experimental drop). Not an LLM: one forward pass reads text + images and returns probabilities for your questions โ no generated tokens, 200-300ms per decision. Billed on input only: $0.04/1M, output $0.00; a 1024x1024 image counts ~1,536 tokens. On six real apps (ticket filtering, code search, web agents, visual inspection) it matches or beats GPT-6.1 Sol on four at 19-200x lower cost, answering significantly faster. Three question types: yes/no, choice, score. API model ID d1; also on OpenRouter and Vercel (text-only there so far). Open weights promised for upcoming models.
Best: Classification-style decisions at scale โ filtering, routing, inspection; cheaper than prompt-hacking an LLM into JSON
CHEAPEST
joey-1 (Baseten)
2B decision model โ $0.01/1M, MIT weights, 20ms P50
In
$0.01
per 1M
Out
$0.00
per 1M
$0.01/1M input, $0 output2B params, MIT on HFJev 0.4: 63.3 โ best <4BP50 20ms end-to-end
Base Labs' third decision model (Oct 29) takes the price floor from Clef-flash's $0.038 straight down to a cent: joey-1 returns a label + calibrated confidence in a single forward pass โ no generated tokens. Baseten-measured P50 end-to-end latency of 20ms on a 2,000-token request, 2.5x faster than 9B elephant-1 (49ms); the family's high-volume first line, with elephant-1 or 27B mammoth-1 as the escalation path. Jev Decision Index 0.4 public split: 63.3 โ 37th of 230 but the top score of any model under 4B (sibling elephant-1 72.2, Clef 61.7). Post-trained from Qwen3.5-2B on 6.21M examples (Brier-calibrated, trained โI don't knowโ abstention), text + image input, full fine-tune on 16 H100s. Model APIs (16K ctx, zero data retention by default), MIT weights live on Hugging Face, fine-tunable on Baseten Training Jobs; 262K ctx on dedicated deployments. Example from Baseten: 10M monthly support tickets (~5B tokens) cost ~$50/month on joey-1 vs $500 on mammoth-1.
Best: Latency-sensitive, high-volume decisions โ ticket triage/routing, spam & phishing, guardrails, LLM-router escalation first line
OPEN
Sarvam Prism
India's decision model โ 105B open weights, 44ms median on one GPU
In
$0.00
per 1M
Out
$0.00
per 1M
In
FREE
local
Out
FREE
local
No generated textSarvam 105B LoRA71.8 @ 44ms (B200 FP8)Apache 2.0 weights
Sarvam AI's entry into the decision-model wave (Nov 3) โ and its first open-weight member. Prism is a LoRA fine-tune of Sarvam 105B: give it a state, a question and fixed options, and a small scoring head returns a probability for each option in a single forward pass โ 'it does not generate text, so the answer is always one of your options'. Vendor-measured across 25 decision models: 71.8 average over three indices at a 44ms median on a B200 โ the highest-scoring model under 100ms in its test set โ after an engineering sprint that took the median from 664ms to 44ms (FP8 conversion, all 128 experts in one grouped kernel, CUDA Graph capture, adapter merged into the weights; 15x faster with no accuracy drop). Calibration pitch: errors stay inside reported confidence bounds and answers are invariant to the order options are presented in. Weights are Apache 2.0 on Hugging Face (sarvamai/prism, ~106GB FP8 on one 120GB GPU โ B200/H200 class) with a hosted API on Sarvam Indus. The category TypeSafe's Jev named (Oct 5) now spans Liquid d1, Microsoft-Decision-1, Cloudflare Clef, SubQ Snap and Baseten joey-1 โ and adds an open-weight, India-hosted entrant at 105B scale.
Best: High-volume enterprise decision workflows you want to self-host โ routing, triage, verification โ at open-weight control
NEW
Liquid LFM3-1.7B
Oct 27 โ 1.7B foundation family: $0.08/$0.20 hosted, open weights
In
$0.08
per 1M
Out
$0.20
per 1M
1.7B foundation128K ctxCached $0.02Weights on HFSingle-GPU deploy
Liquid AI's LFM3 generation (Oct 27): six compact models sharing a 1.7-billion-parameter foundation โ language (LFM3-1.7B), vision (LFM3-VL-2B), audio (LFM3-Audio-2B) and omni (LFM3-Omni-2B) variants around ~2B parameters โ built to move enterprise work onto company-controlled devices: browser agents, voice commands, image recognition, routing, document extraction. Hosted console rates: LFM3-1.7B $0.08/$0.20 per 1M (cache $0.02); VL-2B $0.10/$0.50; Audio-2B and Omni-2B $0.10/$1.00 with audio tokens at $2/$10 per 1M. Weights for 1.7B, VL-2B and Audio-2B are on Hugging Face (LiquidAI org); not yet on OpenRouter at launch. Liquid claims its smallest text model outperforms substantially larger rivals on selected browser-automation and tool-use benchmarks โ company-reported, treat as vendor numbers. One GPU per customized deployment; the non-transformer LFM architecture stays the pitch (efficient sequential data handling at tiny scale).
Best: On-device/edge enterprise agents โ browser automation, voice control, document routing where cloud APIs won't do
NEW
Liquid LFM3-Omni-2B
Tiny omni model โ $0.10/$1.00 hosted, audio in+out
In
$0.10
per 1M
Out
$1.00
per 1M
~2B paramsAudio+image+text inAudio out ($10/1M audio)Cached $0.02Weights on HF
The omni member of Liquid AI's LFM3 family (Oct 27): one ~2B-parameter model that takes text, images and audio in and responds in kind โ no stitched pipelines โ at $0.10/$1.00 per 1M text tokens hosted (audio tokens bill at $2 in / $10 out per 1M audio tokens, not per minute), cache reads $0.02. Sits alongside LFM3-1.7B (text, $0.08/$0.20), LFM3-VL-2B (vision, $0.10/$0.50) and LFM3-Audio-2B in the same 1.7B-foundation generation, all deployable on a single GPU or on-device. Text/VL/Audio weights are open on Hugging Face; Omni listing included in the family announcement. For voice-agent plumbing at this size the closest comparators are MiniMax and Qwen omni tiers at 10x the parameter count โ this is the smallest full-omni stack shipped by a Western lab.
Best: Edge voice agents, on-device assistants, private multimodal pipelines on enterprise hardware
NEW
Step 5 Preview (StepFun)
600B agent MoE โ $1/$2.70, weights on HF (Oct 15)
In
$1.00
per 1M
Out
$2.70
per 1M
1M ctx600B MoE / 27B activeAA 44 โ #12Open weights on HF โ Oct 15
StepFun's flagship agentic model (launched ~Sep 18 on StepFun's own API at the same $1/$2.70; hit OpenRouter Oct 8 as a router). 600B-param sparse MoE with 27B active, 1M-token context, text+image+video input (max output 64K on StepFun's docs; 128K on OpenRouter; reasoning + tool calling). StepFun's own pitch: 'the Pareto frontier marks the best trade-offs between intelligence and cost โ progress begins when that boundary shifts outward.' AA Intelligence Index v4.3.2: 44 โ #12, above Kimi K3 and GLM-5.3, at ~$1.03 per Index task vs Sonnet 5.5's $1.08โ$7.60 by effort. Vendor numbers: DeepSWE v1.1 67.7% (vs Kimi K3's 67.5), Terminal-Bench v2.1 85.0%, GPQA Diamond 93.5%, BrowseComp 88.7%, FrontierFinance 66.4% โ beating GPT-6 Astra's 55.0; a weaker Terminal-Bench v4 of 33.3% vs GLM-5.3's 41.9 shows the aggregate sits below the highlights. Very verbose: 160M output tokens across the AA run, roughly double the median. The stand-out demo: 24 hours unsupervised on an H100 optimizing an MLA GPU kernel to 508 TFLOPS (vs Opus 5's 493). Cache hits bill at ~10% of input ($0.10) โ effective blended rate ~$0.54/1M. The promised Oct 15 open-weights drop did land on StepFun's own Hugging Face org: stepfun-ai/Step-5-Preview plus BF16 and NVFP4 variants are live (created early Sep, repos updated Oct 15; licence tag 'other' โ check the custom licence terms before commercial serving; earlier reporting that no vendor repo existed appears to have missed it). MIT-style licence still claimed for the eventual full open-source release.
Best: Agentic coding, finance/research agents, long-horizon work at mid-tier prices โ strongest sub-$5 open-weights candidate since MiMo-V2.6
NEW
MiMo-V2.6-Pro
New open-weight king โ 1.02T MoE, omni-modal, $0.435/$0.87
In
$0.43
per 1M
Out
$0.87
per 1M
1M ctx (128K out)1.02T MoE / 42B activeNative omni-modalAA 46.32 โ #1 openMIT weights
Xiaomi's new flagship (Sep 22), released and open-sourced together: 1.02T-param MoE (42B active) with hybrid attention and multi-token prediction, trained via massively scaled RL on the recursive-self-improvement path. AA Intelligence Index v4.3.2: 46.32 (max) โ the strongest open-weight model measured, ahead of GLM-5.3 and Kimi K3 (~44) and past Grok 4.7's 46 on the same revision. Native omni-modal: text, image, video and audio input; 1M context, 128K output. API: $0.435/$0.87 per 1M with cache hits at $0.0036 โ flat vs V2.5, which Xiaomi pegs at 1/20 to 1/60 of overseas pricing at equal intelligence. MIT weights, tech report, Distill-Qwen-9B and RL training resources on Hugging Face; an ultraspeed variant serves the same model faster.
Best: Agentic coding, multimodal understanding, open-weight frontier work at 1/20 of US frontier pricing
NEW
MiMo-V2.6-Flash
310B open MoE at $0.14/$0.28 โ omni-modal budget king
In
$0.14
per 1M
Out
$0.28
per 1M
1M ctx310B MoE / 15B activeOmni-modalCache $0.0028MIT weights
The MiMo-V2.6 high-volume tier (Sep 22): ~310B-param MoE with ~15B active, native omni-modal input, 1M context. $0.14/$0.28 per 1M with cache hits at $0.0028 โ among the cheapest multimodal APIs available, MIT-licensed with full weights open. Shares the V2.6 series pricing with Pro and the ultraspeed variant; Token Plan subscriptions cover the whole suite.
Best: High-volume multimodal agents, cost-sensitive production, local-adjacent open weights
BUDGET
Solar Mini 4
Upstage's 3B-active agent MoE โ $0.05/$0.20 promo
In
$0.05
per 1M
Out
$0.20
per 1M
524K ctx35B MoE / 3B activeCache $0.01Promo 50% off
Upstage's compact agentic model (Sep 23): 35B-param MoE with 3B active and a 524K context window. $0.05/$0.20 per 1M on OpenRouter โ a 50% promotional discount off the $0.10/$0.40 standard rate; cache reads $0.01. Korean-language strength plus English and Japanese. No published frontier-benchmark ceiling โ it's built for throughput, not leaderboards.
Best: High-throughput agentic pipelines, document retrieval, long-conversation agents
UPDATED
DeepSeek V4.1-Flash
Checkpoint 1017 (Oct 17): AA 45 vs 39.5 โ +5.5 at same price
In
$0.15
per 1M
Out
$0.60
per 1M
1M ctx552B MoE (8B prefill / 16B decode)AA 45 (39.5 pre-1017)DeepSWE 77.0MIT weightsCache $0.003
Checkpoint replacement shipped Oct 17, 2026 (558 points on HN): DeepSeek-V4.1-Flash-1017 swaps into the same release โ same 552B-param CED backbone (8B active prefill, 16B decode), native vision, 1M context, 384K max output, ~510GB FP8, MIT licence โ and the API model name deepseek-flash already serves it at unchanged pricing: $0.15/$0.60 off-peak, $0.30/$1.20 peak, cache hits $0.003/$0.006. Model-card gains: DeepSWE v1.1 74.2 โ 77.0 (vendor table puts it above Claude Opus 5's 74.0), Terminal-Bench 4.0 31.2 โ 42.3, Automation-Bench 54.8 โ 60.5, HLE w/o tools 36.8 โ 42.6, Agents' Last Exam 31.8 โ 35.4, with shorter outputs and faster task completion. Now independently measured: AA Intelligence Index v4.3.2 scores the 1017 checkpoint at 45 (vs 39.5 for the September weights โ +5.5 on in-place weights at the same price; index-level confirmation of the agent gains; per-eval AA tables use the Max variant). Same day: Search and Fetch web-retrieval APIs opened to all users at $0.75 per 1,000 successful requests (5 QPS cap; server-side calls via DeepSeek Harness/Claude Code/Codex free until Nov 1).
Best: Cache-heavy agent loops, coding agents, high-volume multimodal โ cheapest 1M-context frontier-adjacent API
NEW
GLM-5.3
Z.ai's open-weight coder โ now #2 open-weight
In
$1.40
per 1M
Out
$4.40
per 1M
743B paramsOpen weight (pending)CodingGLM Coding Plan & ZCode
Z.ai (formerly Zhipu AI) calls GLM-5.3 the most capable open-weights model for coding. 743B-param model built by scaling post-training on the GLM-5.2 base. Weights are now public on Hugging Face (MIT-style GLM-5.3 licence, Aug 27) โ 51 quantizations available. Live via GLM Coding Plan, ZCode and the API. AA Intelligence Index v4.3 (max): 44.9 โ the strongest open-weight model measured. GLM-5.3-Flash (41.9) is the cheap sibling at $0.15/$0.50 with 1M ctx after its Sep 9 promo ended. 'Dramatic improvement over GLM-5.2 with fewer output tokens.' GLM-5.2 API was ~$1.40/$4.40 โ roughly a tenth of US frontier per-token rates. Sep 22: Xiaomi's MiMo-V2.6-Pro (46.3 on AA v4.3.2) takes the open-weight #1 spot; GLM-5.3's 44.9 (v4.3) is now the strongest Z.ai model.
Best: Agentic coding, open-weight workflows, cost-sensitive enterprise
NEW
GLM-5.3-FlashX
New $0.37/$1.25 mid tier
In
$0.37
per 1M
Out
$1.25
per 1M
1M ctxFlash-FlashX tier (Z.ai)Cache $0.075
Z.ai's newest tier between GLM-5.3-Flash ($0.15/$0.50) and the GLM-5.3 flagship ($1.40/$4.40): $0.37/$1.25 per 1M, cache $0.075, 1M ctx. The FlashX-Legal legal-domain variant (Nov 3 on OpenRouter) prices identically โ see its own card. FlashX targets the same cheap-but-stronger-than-Flash agentic slot K2.7 Code just took at Moonshot. Note the Flash tier briefly whipsawed to $0.50 output before settling back on Z.ai's official list ($0.15/$0.50).
Best: Budget agentic coding between Flash and flagship
NEW
GLM-5.3-FlashX-Legal
NEW โ legal-domain FlashX at the same $0.37/$1.25
In
$0.37
per 1M
Out
$1.25
per 1M
1M ctxLegal mid+post-trainingMandatory thinkingMCP tools ready
Z.ai's legal specialist (Nov 2 changelog / Nov 3 OpenRouter): GLM-5.3-FlashX post-trained with legal-domain mid-training and reinforcement learning in task environments modelled on real legal workflows โ document understanding, contract review, cross-document consistency checks, legal research and drafting/revising documents. Designed to run in agent setups wired to legal data sources and document-editing tools (MCP services). Keeps the FlashX base: 1M ctx, mandatory thinking mode, function calling, structured output, streaming โ at identical pricing: $0.37/$1.25 per 1M, cache $0.075. The third vertical to get a dedicated cheap tier this autumn, after cyber (Fairwind Program) and life-sciences (Rosalind).
Best: Contract review, legal research, document drafting, cross-document checks
NEW
DeepSeek V4-Pro
Still live โ V4.1-Flash now beats it on cost & speed
In
$0.66
per 1M
Out
$1.98
per 1M
128K ctxGA Aug 13DeepSWE 62.7%Terminal-Bench 87.9Off-peak price
DeepSeek's prior flagship out of preview (Aug 13). DeepSWE 62.7%, Terminal-Bench 2.1: 87.9. AA Intelligence Index v4.3 (max): 36.3. Peak/off-peak billing: off-peak $0.66/$1.98, peak $1.32/$3.96. Update Sep 10: DeepSeek's own tests put V4.1-Flash ahead on performance, cost, speed and runtime. From Sep 14 new deepseek-v4-pro requests can still be served (DeepSeek walked back the retirement after user demand) but check whether V4.1-Flash at $0.15/$0.60 off-peak does the same job for less. Migrate to DeepSeek V4.1-Flash unless you specifically need V4-Pro's HLE lead (42.7 vs 36.8).
Best: Autonomous agents, software engineering, cost-sensitive high-volume workloads
NEW
Qwen3.8-Omni-Flash
Omni-modal agents โ audio+video in one model, 1/5 of Gemini's price
In
$0.15
per 1M
Out
$0.47
per 1M
1M ctx (991K in / 131K out)Audio + video native262K reasoningAPI-only, no weightsCache $0.016
Alibaba's first omni-modal agent model (Sep 18). Accepts text, images, audio and video natively โ up to 3 hours of audio, 2-hour video files โ and returns text, with thinking on by default, function calling and web search. Built on the open-weight Qwen3.8-Flash-Next architecture but served API-only (QwenCloud, Model Studio); no downloadable weights. QwenCloud pricing: $0.15/$0.47 per 1M, implicit cache hits $0.016 โ audio input under $0.01/hour, ~$0.20 per minute of 720p video. +25% average across 29 evals over Qwen3.5-Omni-Plus; WildClawBench-MM 34.5 to 71.0, UniClawBench 69.6 (Gemini 3.8 Flash: 69.0), though Gemini still leads AgenticVBench (45.0 vs 36.8). SWE-bench Pro 63.3, LiveCodeBench v6 92.6. Qwen-MM-Plugins (Apache-2.0) add audio/video handling to Claude Code, Gemini CLI and Qwen Code; a Realtime variant handles live audio.
Best: Meeting/call analysis, video research agents, media indexing โ audio+video agentic work at a fifth of Gemini's price
NEW
Qwen 3.8-Max
Alibaba's flagship โ 'second only to Fable 5'
In
$2.00
per 1M
Out
$6.00
per 1M
1M ctx (flat tier)2.4T MoE / 95B activeMultimodalOpen weights soonAA Index 58 pts
Alibaba's most capable model. 2.4T-param MoE with 95B active. $2/$6 per 1M โ flat across the entire 1M-token context window (no long-prompt surcharge). Cached input at $0.25/1M. Claims 'second only to Fable 5'. AA Intelligence Index v4.1.1: 58 pts (#8 globally, between GPT-5.6 Sol high and GPT-5.6 Terra max). Multimodal (text + visual). Open weights promised within days of GA. OpenAI and Anthropic API compatible.
Best: Long-context work, coding, enterprise โ flat pricing across 1M context
RETIRED
MiniMax M2
Retired Nov 18, 2026 โ replacement M3.1 Flash
In
$0.30
per 1M
Out
$1.20
per 1M
Retired Nov 18Was $0.30/$1.20Open weightSuperseded
Retired: MiniMax retired the M2.5/M2.7 families Nov 18, 2026 (new users cut off Oct 19; PAYG calls stop at retirement; M Plan/Token Plan already routed to M3.1 Flash since Oct 19). The original M2 agent model โ $0.30/$1.20 per 1M at ~100 TPS, ~8% of Sonnet's price โ rides the same sunset. Replace with MiniMax-M3.1-Flash: 42.7 on AA v4.3.2 at $0.15/$0.60 (half the price of M2), 428B/23B-active open weights since Oct 23, 1M ctx, adaptive thinking.
Best: Agents, coding, tool use, cost-sensitive agentic workflows
NEW
Tencent Hy3
Global open API โ cheapest per-token
In
$0.13
per 1M
Out
$0.53
per 1M
256K ctx295B MoE / 21B activeApache licenseOpen weights
Tencent's reasoning and agent model (formerly Hunyuan). 295B MoE with 21B active per token. Apache-licensed weights on HuggingFace. OpenRouter from $0.13/$0.53 per 1M โ among the cheapest open APIs available. Topped OpenRouter usage leaderboard within a week of its Jul 6 launch. Global access via WorkBuddy (free until Aug 31), Tencent Cloud TokenHub, and API.
Best: Reasoning, agent tasks, cost-sensitive high-volume apps
NEW
Tencent Hy4
Final with vision + computer use โ frontier-class under $1 input
In
$0.83
per 1M
Out
$2.50
per 1M
770B-class MoEText+image in โ visionCLI + GUI computer use1M+ ctxOpen weights promised
Tencent's generational leap went final (Nov 4): Hy4 is the last version of the Aug 28 preview โ official announcement โ and now 'ranks among the world's leading open-source AI models' per Tencent, with two firsts: native visual understanding (fine-grained image processing, complex chart analysis, mathematical visual reasoning) and computer use across both command-line and graphical interfaces, with specialized training on professional Blender and CAD software for long-horizon, sophisticated tasks. Tencent's own tables put APEX-Agents v1.1 at 75.2 โ edging Claude Opus 5.5's 74.8 (vendor-measured) โ and WorkBuddy evals average 83.18 with a 96.15% pass rate, first place in 9 of 12 mainland-China scenarios. Pricing unchanged from preview: $0.834/$2.501 per 1M ($0.042 cache hits) via Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's Dayuan and ima, plus OpenRouter and Hermes/OpenCode โ OpenRouter currently lists the route at half list ($0.417/$1.25) under a discount field of 0.5. Launch window: TokenHub 50% off through Dec 31; WorkBuddy and CodeBuddy free Nov 4โ15. Weights + system card promised 'in the coming days' (cybersecurity and bio-safety focus) โ the Hy4 Preview pattern was Apache 2.0 on Hugging Face; Hy4 Flash is teased as the cheaper tier.
Best: Long-document work, coding, office productivity at open-weight prices
NEW
Kimi K3.1 Pro
#4 overall โ a third of Sonnet's price
In
$2.00
per 1M
Out
$10.00
per 1M
1M ctxMultimodalMoE 2.8T / 104B activeAA v4.3.2: 55 (#4)Terminal-Bench 4.0 top scoreWEIGHTS LIVE Nov 2 โ MXFP4
Moonshot's K3.1-series debut (Oct 23): 2.8T-total / 104B-active MoE scores 55 on AA v4.3.2 โ #4 overall, behind only Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56), and ahead of GPT-6 Astra, Fable 5.1 and Gemini 4 Argon. Highest Terminal-Bench 4.0 score measured: 67.2% on AA's harness (Moonshot's own Claude-Code run: 65.7% vs Astra 58.2%, Opus 5.5 66.4%). GDPval-AA 1,846 Elo (3rd, +312 over K3), ~74 tok/s, $2.11 per AA task. Kimi API: $2/$10 per 1M (cache $0.20; K3.1 Pro Highspeed $4/$20, 2โ3x faster; 1M ctx for Pro-tier members). Hallucinations down 65% on Moonshot's finance stress test (8.3% vs K3's 23.4%). Cyber-offense data deliberately pulled from the open release (in-house exploit-success 90.5% โ 81%) โ an open-weights precedent under scrutiny. Weights LANDED Nov 2 on Hugging Face (moonshotai/Kimi-K3.1-Pro) โ two days ahead of Moonshot's committed Nov 4 date: 2.8T/104B-active open weights with MXFP4 weights / MXFP8 activations (quantization-aware training), a 401M-param MoonViT-V2 vision encoder, and Gated DSA sparse attention replacing K3's full-attention MLA layers for efficient 1M-token serving. Heft warning โ big-box open weights again: plan ~600GB-class VRAM or quantized serving. It is now the best open-weight model ever measured (AA 55, past MiMo-V2.6-Pro's 46.3), and the cyber-offense data stays stripped from the release. K3 at $3/$15 remains the budget 3T-class fallback. K3 at $3/$15 remains the budget 3T-class fallback.
Best: Frontier coding/agents at open-weight prices; agentic work, long-horizon knowledge work
SUPERSEDED
Kimi K3
Superseded โ K3.1 Pro scores 55 (#4)
In
$3.00
per 1M
Out
$15.00
per 1M
1M ctxMultimodalMoE 2.8T / 104B activeAA v4.3.2: 43.6Superseded by K3.1 Pro
Moonshot's flagship โ now superseded by Kimi K3.1 Pro (Oct 23), which scores AA v4.3.2 55 (#4 overall, +11) at $2/$10 and takes the Terminal-Bench 4.0 crown; K3.1 Pro weights landed Nov 2 (MXFP4). K3 itself: 2.8T-param MoE, world's largest open-weight model, 57 on AA v4.3 at launch (later 43.6 on v4.3.2), native multimodal, 1M ctx, $3/$15, weights public Jul 27 (1.56TB). Still the cheapest 3T-class weights you can download today.
Best: Cheapest 3T-class open weights; K3.1 Pro for new builds
NEW
Kimi K2.6
Repriced โ was $0.60/$2.50
In
$0.95
per 1M
Out
$4.00
per 1M
256K ctxOpen weightMoE 1T/32BCache $0.16
Moonshot's official table moved K2.6 to $0.95/$4.00 per 1M (Sep era: $0.60/$2.50) โ cache reads $0.16. 1T params, 32B active, 256K ctx. Beats GPT-5.4 and Opus 4.6 on SWE-Bench Pro; K2.7 Code joins it at the same $0.95/$4 (Highspeed $1.90/$8). Still far under US frontier rates, but the '88% cheaper than Opus' line is dead.
Best: Coding, agents, long-horizon tasks
CHEAPEST
Qwen 3.6 Plus
1M context, free tier
In
$0.10
per 1M
Out
$0.30
per 1M
1M ctxReasoningFree tier
Alibaba's latest. Mandatory chain-of-thought reasoning. Free tier available. Topped 6 coding benchmarks on release.
Best: Budget coding, massive context
NEW
Llama 4 Scout
10M context MoE
In
$0.15
per 1M
Out
$0.55
per 1M
10M ctxOpen weightMoE 109B
Longest context of any open model. 109B total, 17B active. Multimodal. Runs on 24GB VRAM.
Best: Massive context, multimodal, local
Llama 4 Maverick
Frontier coding MoE
In
$0.20
per 1M
Out
$0.80
per 1M
1M ctxOpen weightMoE 400B
Beats GPT-4o on coding. 400B total, 17B active. 128 experts. Frontier quality at MoE prices.
Best: Coding, complex reasoning
DeepSeek V3.2
Matches GPT-4o
In
$0.27
per 1M
Out
$1.10
per 1M
128K ctxOpen weightMoE 685B
94.2% MMLU matching GPT-4o. 685B MoE with 37B active. Best open model for general knowledge.
Best: General knowledge, research
FREE
Qwen3-Coder 8B
Local coding king
In
$0.00
per 1M
Out
$0.00
per 1M
In
FREE
local
Out
FREE
local
32K ctxLocal only8B dense
Runs on any 8GB GPU. 92 programming languages. 80-150 tok/s. Best local coding model under 10B. Set it up locally →
Best: Local coding, autocomplete
FREE
DeepSeek R1 Distill 14B
Local reasoning
In
$0.00
per 1M
Out
$0.00
per 1M
In
FREE
local
Out
FREE
local
Local onlyReasoning10GB VRAM
Chain-of-thought reasoning on 10GB VRAM. The sweet spot for local reasoning. 55 tok/s on modern GPUs. Run it offline →
Best: Local reasoning, budget hardware
๐ก Did You Know?
GPT-6 Cyber โ a $20/$100 row most approved accounts can't call
The pricing guide's first Critical-rated entry breaks the usual shape: a live price row โ $20/$100 per 1M token, 2x GPT-6 Astra, the most expensive OpenAI line โ gating a model that only selected Daybreak Red customers can call. The capability jump is real: ExploitGym 48.5% (GPT-5.6 Cyber: 32.1%), RemoteExploitBench 95% completion where Astra under Daybreak Blue refused every task, black-box web-pentest recall 40.3%, plus fielded results โ 948 Linux kernel privilege-escalation exploits (vs 166 for GPT-5.6 Sol, same harness) and 81 novel high-severity Firefox vulnerabilities with fixes Mozilla shipped. OpenAI's own system card supplies the counterweight: on the 288-task Irregular FrontierCyber set Astra solved more (96 vs 90), and neither touched the 15 hardest tasks. The interesting work was alignment: honeypot out-of-scope exploitation fell from 5.51% to 0.07% because scope-respect was trained as a first-class objective โ a model this good at attacking has to stay inside the rules of engagement. Continues the autumn pattern: price row first, permission list in between, general access never promised with a date.
Sarvam Prism โ the decision-model wave goes Indian and open-weight
The category TypeSafe named after Kahneman's System One (Oct 5) picked up its seventh member and most unusual lineage yet on Nov 3: Sarvam Prism, from India's Sarvam AI โ a LoRA fine-tune of its own Sarvam 105B that reads a state, question and fixed options in one forward pass and returns a probability per option, never generating text. Sarvam's evaluation across 25 decision models reports 71.8 average over three indices at a 44ms median on a single B200 โ the top score under 100ms in its test set โ after an engineering sprint that took the median from 664ms to 44ms: FP8 conversion, all 128 experts in one grouped kernel, the forward pass captured as a CUDA Graph, and the fine-tuning adapter merged into the weights. Calibration claims: errors within confidence bounds, answers invariant to option order. Weights are Apache 2.0 on Hugging Face (~106GB FP8, one 120GB GPU), with a hosted API on Sarvam Indus. A week after Baseten's joey-1 dropped the price floor to $0.01/1M, the open-weight decision model arrives โ enterprise calibration you can fine-tune on your own data and host on your own GPU.
Baseten joey-1 โ the first decision model at a penny
Baseten's price floor on Oct 29 isn't an LLM: joey-1 is a 2B decision model at $0.01/1M (cached and output $0.00), a quarter of the previous floor (Clef-flash $0.038, set Oct 9) and a quarter of Liquid d1 ($0.04). One forward pass returns a label and a calibrated confidence score; Baseten measured P50 end-to-end latency of 20ms on 2,000 tokens โ 2.5x faster than its 9B sibling elephant-1, 5x faster, it says, than comparable decision models. Its Jev Decision Index 0.4 public split score of 63.3 places it 37th of 230 overall but the top score of any model under 4B โ the family (mammoth-1 27B $0.10, elephant-1 9B $0.035) shares one API: one-line switch, escalation path preserved. Post-trained from Qwen3.5-2B on 6.21M decision examples with a trained โI don't knowโ output, MIT licence, weights on Hugging Face. The category is weeks old (d1 opened it Oct 5) and its price floor has already fallen 4x โ the classic commodity curve arriving in fast-forward.
Gemini 4 Carbon โ the first model to top the v5 leaderboard
Google's Oct 28 announcement landed one day after Artificial Analysis rebased its Intelligence Index to v5.0 (Oct 27 โ 60% private test sets, Cognition's FrontierCode 2.0 and Terminal-Bench-Science in, GDPval/SciCode/LCR out), and Carbon immediately took #1 at 59 on 4 out of 4 intelligence units. The rebase matters: v4.3.2 numbers (Opus 5.5's 58, Sonnet 5.5's 56, Kimi's 55) don't transfer โ on v5 Opus 5.5 is 58.4, Astra 56.9, Sol 54.5, Sonnet 54.2, Terra 51, and Argon (high) 50.4. The open-weights story inverted: in v4.3.2 Kimi K3.1 Pro sat 3 points from #1; on v5 the gap is 12 points, and among models with weights actually released, DeepSeek 1017 and GLM-5.3 tie at 32 โ Claude Haiku 5.5 (36.2) outscores every released open-weight model. Carbon's own claims are cyber-shaped (SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0%) and priced like Argon ($2/$10 intro, then $4/$20) โ but it's Fairwind-gated first, the same play that delayed Argon 13 days from announce to GA. Meanwhile the price-performance leader of the v5 top ten is GPT-6.1 Terra at $0.96/task โ a tenth of Opus 5.5 max's $10.04.
Liquid LFM3 โ the two-billion-parameter counter-program
While the frontier labs chased bigger context windows, Liquid AI shipped the opposite bet (Oct 27): LFM3, six compact models sharing a 1.7-billion-parameter foundation โ text (LFM3-1.7B, $0.08/$0.20 per 1M hosted, cache $0.02), vision (LFM3-VL-2B, $0.10/$0.50), audio (LFM3-Audio-2B) and omni (LFM3-Omni-2B, $0.10/$1.00 with audio tokens at $2/$10 per 1M) โ sized to run customized on a single GPU or entirely on company-controlled devices. The pitch: browser agents, voice commands and document extraction that never leave the enterprise's own hardware. Weights for the 1.7B, VL-2B and Audio-2B are on Hugging Face; Liquid's claim that the smallest text model beats substantially larger rivals on selected browser-automation and tool-use tasks is company-reported. The same day OpenAI quietly surfaced a gpt-6.1-terra-pro serving variant at Terra's $1/$5 rates (pro-mode reasoning burns several-times more tokens) and Mistral Large 4's promised Oct 27 weights slipped again โ a week when the small-model floor kept dropping while the big-weight drops kept not landing.
GPT-6.1 Terra โ OpenAI halves the workhorse price (again)
GPT-6.1 Terra launched Oct 26 as the third member of the 6.1 family, and the story is the price curve: $1/$5 per 1M โ half of GPT-6.1 Sol, half of GPT-5.6 Terra's own $2/$12, and one-tenth of GPT-6 Astra's $10/$50 โ with cached input at $0.05, batch $0.50/$2.50 and Fast mode $2/$10, all exactly half of Sol's equivalents. The pitch is the office-mill: visual reasoning plus precise computer use โ match product photos to catalog entries, interpret technical drawings, act on application screens. Official numbers: Agents' Last Exam +15 pts over 5.6 Terra at medium effort for ~a third of the cost, on par with Astra at a tenth of the price; at max effort, over 3x the token efficiency of Opus 5.5 at a quarter of its per-token price, ~9x lower cost per task; Terminal-Bench 4.0 50.3% (5.6 Terra: 21.5%); error rate 4.3% at max on OpenAI's internal error-inducing benchmark, down from 53.6%. Artificial Analysis's first measurement (Oct 27): 48 on v4.3.2 max โ six under Sol's 52, a dollar less; $0.40 per Index task on 130M tokens. Live in ChatGPT Work and Codex for paid plans (release notes Oct 26 note some Power presets swapping Terra at Medium in for Sol at Light) and on the API as gpt-6.1-terra. 6.1 family so far: Sol (Sep 29, near-Astra at a fifth of Astra), Terra (the workhorse), and the Astra upgrade OpenAI cancelled rather than ship.
Mercury 3 โ diffusion finally scales
Inception's Mercury 3 (Oct 19) is the answer to whether diffusion LLMs can compete: the largest diffusion model ever released is more than 40x the size of Mercury 1, holds speed at 1,068 tok/s on commodity NVIDIA GPUs, and stretches context from 260K to 1M tokens. The vendor claim is on par or better than GPT-6 Luna and Claude Haiku 5.5 across a range of benchmarks at 4โ8x their speed; the sharpest demo is production web agents, where Mercury 3 completed 50% of 50 live Mind2Web tasks, ahead of GPT-6.1 Sol, at the same token volume and step count. Independent AA Index data is still pending (Mercury 2.5 measures 12), so treat quality claims as vendor-measured for now. Price: $0.20/$0.70 list, 50% launch discount at $0.10/$0.35, cached input $0.01 โ live on the Inception API, OpenRouter, Baseten and Vercel with 100M free tokens on new keys. Voice, tutoring and browser agents now have a non-autoregressive option at Haiku-class money.
MiniMax M3.1 Flash โ Space Bunny comes in from the cold
The anonymous 'Space Bunny' โ the leaderboard phantom that processed 200+ trillion tokens on llama.cpp without ever identifying itself โ turned out to be an early version of MiniMax M3.1 Flash (Oct 19): $0.15/$0.60 per 1M with cache hits at $0.015, 70% cheaper to run than M2, ~3x faster inference, 1M context (524K output), text+image in. MiniMax's tables put it at the leading group on AutomationBench-AA, with BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7 and Agents' Last Exam 29.6 โ first among Chinese models on both โ plus CyberGym 82% for vulnerability work. Independently measured within a day: 42.7 on the AA Intelligence Index, #21 of 351, at $0.22 per task โ a point behind GPT-5.6 Terra for a ninth of its price, with very high verbosity as the catch (180M tokens across AA's run, ~80% above median). The free OpenRouter window closed Oct 22 at 12:00 UTC (the :free variant retired); M3.1-Flash open weights landed Oct 23 as promised โ MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face, ~428B total / 23B active under the minimax-community licence, with sparse attention, DSpark speculative decoding and a published vLLM recipe โ and Pro and Flash Pro are teased as the intelligence tier to come.
DeepSeek 1017 โ +5.5 intelligence points with the price unchanged
The cheapest way to get smarter this month arrived as a checkpoint, not a model. On Oct 17 DeepSeek swapped the V4.1-Flash weights in place (552B-param CED backbone, 1M ctx, MIT licence) and re-pointed the same deepseek-flash API name at the new checkpoint โ rates untouched at $0.15/$0.60 off-peak, $0.003 cache hits. The Friday HN thread hit 558 points, and the interesting part is what showed up four days later: Artificial Analysis's v4.3.2 index scored the 1017 checkpoint at 45, up from 39.5 for the September weights โ a 5.5-point independent jump on identical hardware-class infrastructure at an identical price, with AA-level confirmation of the vendor's agent gains (DeepSWE 74.2 โ 77.0). For the $1-per-weekend-projects crowd it keeps the title of cheapest 1M-context frontier-adjacent API; for everyone else it's a case study in why checkpoint releases may be the quiet pricing lever of 2026 โ intelligence inflation, repriced for free.
Muse Spark Open โ Meta finally ships open weights, tuned for your desk
Oct 15 delivered the release Meta watchers had stopped expecting: Muse Spark Open, the company's first open-weight model โ 816B total/42B-active MoE on the Muse Spark 1.3 base with extra safety post-training for standalone use, 1M context, text+image+video+audio input. HN's front page (397 points) came not from the benchmark table but the deployment story: co-tuned with Apple, it runs entirely on two Mac Studio M5 Ultras via the latest MLX โ Apple's 512GB-per-box unified memory combines into a 1TB desktop supercomputer at 1.2TB/s, and Meta demoed it driving Muse Code to ship real iOS changes to Ice Cubes and editing DaVinci Resolve footage frame by frame, fully on-device. Weights are free on Hugging Face under the Muse Model License (free for most; model-serving businesses past $20M aggregate revenue must notify Meta and sign before serving). Hosted access prices it exactly like closed Spark 1.3: $1.25/$4.25 (cache $0.15) across OpenRouter, Together, Fireworks, Groq, Modal, Nebius and Baseten. The honest catch sits in Meta's own model card: AA Intelligence Index ~39 โ roughly seven points behind Xiaomi's independently measured MiMo-V2.6-Pro (46.3), and the full eval report is still pending. 'Most capable US open-weight model to date' is a claim about the US-vs-China open field (Meta says it outperforms GLM 5.3 and Kimi K3 on agentic and visual reasoning), not about the AA leaderboard โ and Bloomberg notes Meta kept its genuinely frontier tech (Muse Wonder, 52/56) closed. Fourteen months after the Superintelligence Labs rebuild, Meta's open-weight flag is planted โ smaller than the flagship, priced like the old Spark, built to run on a desk.
Muse Wonder โ Meta finally ships a frontier model
Fourteen months after gutting its AI stack (and its safety teams) under Alexandr Wang, Meta released its first frontier models on Oct 12, trained at the gigawatt-scale Prometheus datacentre in Ohio: Muse Wonder at $4/$20 per 1M ($0.20 cached input) with 1M context and text+image+speech+video input, and Muse Wonder Full, which scores 56 on the Artificial Analysis Intelligence Index โ third-highest measured, two behind Claude Opus 5.5, tied-ish with Sonnet 5.5 (55.8 vs 56.0 unrounded) โ but is unreachable by API: Wonder Full (max reasoning) is sold only through the Muse Code coding agent's $50/$200 subscriber plans (roughly $60/$300 per-1M-equivalent per CNBC). The API-facing Wonder scores 52 โ one point behind Fable 5.1 and GPT-6 Astra โ while running ~6x Sol's cost per task ($4.52 vs $0.72) on 83 tok/s, and the memory story is Meta-flavoured: Active Memory and Active Focus are the pitch, and MemoryBench (85.5-85.6 vs Fable's 79.0, Astra's 81.4) is a Meta-internal benchmark. TechCrunch noted Wonder Full also beats Opus 5.5 on deep legal work (Harvey LAB-AA 13.1 vs 4.2) and that Meta's benchmark table includes scores from Artificial Analysis, Vals AI and Datacurve โ not just its own charts. The frontier has a fourth lab; the bill for it is a subscription, not a token meter.
SubQ Snap โ decision models keep getting weirder
Oct 12 brought a fourth decision-model entrant in a week: SubQ Snap from Subquadratic, a 'long-context decision model' that scores a fixed set of allowed answers rather than generating text โ 384K tokens of text+images per call, $0.05 per 1M input tokens, output free, sub-100-second claims on requests up to 100K tokens, built on the SSA subquadratic architecture the company teased earlier in the year. A 100K-token decision call costs about half a cent of input. It's in public preview at subq.ai on a first-come basis (waitlist when capacity fills), and its Decision Index 0.2.1 61.7 vs TypeSafe Jev's 57.9 is a vendor-measured win โ third-party benchmarking still pending. One week in, the no-generated-tokens category has Liquid d1 ($0.04), Cloudflare Clef-flash ($0.038) and Clef-omni ($0.15, audio+video), Microsoft-Decision-1 ($0.042) and SubQ Snap ($0.05) โ all input-only pricing. Agent plumbing is now its own product category, and it has a price floor under a nickel.
Cloudflare shipped three decision models in eight days
The oddest launch cycle of October belongs to a CDN company: Cloudflare decided on a Friday evening to enter Liquid d1's decision-model space, trained Clef over the weekend, and shipped it the following Thursday (Oct 1-2). Eight days later (Oct 9) the Clef line had tripled: Clef-omni takes audio, video, images and text in one pass โ the first multimodal decision model in a category that was text-only since TypeSafe's Jev โ at $0.15/1M input-only; Clef-flash was cut 58% to $0.038/1M, cheaper than Jev and undercutting d1's $0.04; and Clef dropped from 262ms to 152ms median latency after a move to SGLang. Clef-omni is built on the Qwen3-Omni-30B-A3B-Instruct backbone with the TTS head removed, prefill-scores every option across all modalities at once, and returns calibrated confidence scores through a schema-bound lexical grammar โ a 21-second video with sound scores in ~1.5 seconds. All three are Apache 2.0 on Hugging Face (Cloudflare/clef-omni), so you can self-host past the 64K (24K for flash) hosted context. No output tokens means no output bill โ classification plumbing priced like a utility.
Microsoft-Decision-1 โ the decision-model category goes mainstream
Microsoft shipped Microsoft-Decision-1 on Oct 9: not an LLM but a decision model that reads a brief and returns probabilities over predefined choices โ continue, stop, retry, escalate, route โ at $0.042 per 1M input tokens with output free. Microsoft's 36-benchmark suite (~150K questions, kept blind from training) puts it first on accuracy and fastest measured: 4.5x the runner-up Quyet-1.0-Large, 35x faster than GPT-6 Sol, classifying 1M texts for ~$11 where GPT-6 Sol costs ~$2,434. The base is telling: Qwen3.5-9B plus single-pass decision-scoring post-training, with rebases on MAI and OpenAI models planned โ the HN thread notes this is the third decision model in a fortnight built on a Qwen base (Cloudflare's Clef, Strands' decider, now Microsoft's). Internal results: Xbox Research categorized 10K+ gaming feedback items at GPT-6-Sol-comparable quality 14x faster at 1/200 the cost. One week after Liquid d1's $0.04/1M input-only pricing, the biggest software company on earth validated the category. The agent-economy plumbing layer is now its own product line.
Haiku 5.5 โ the small-model floor drops another 90%
Claude Haiku 5.5 (Oct 8 NZ / Oct 7 US) rewrites the small-model math: $0.10/$0.50 per 1M for prompts up to 100K โ a tenth of Haiku 4.5's $1/$5, with 1M context (long-context rates $0.50/$2.50 above 100K), cache reads from $0.01, and 300K output on the Batch beta. Anthropic pegs average running cost ~75% below Haiku 4.5 after accounting for the new tokenizer's ~30% token inflation. AA's first full Index measurement (published Oct 8): 43.4 at max โ inside a point of Kimi K3's 43.6 and #14 of 687 measured models, at roughly a tenth of Kimi's per-token price. The capability leap is the story: OSWorld 2.1 72.4% (4.5: 15.7%), Terminal-Bench 4.0 39.2% (4.5: 0.0%). Paired with Sonnet 5.5's cache reads halved to $0.10 the same day, Anthropic is pricing for the subagent economy: big model plans, cheap models swarm.
Step 5 Preview โ StepFun moves the Pareto frontier
StepFun's Step 5 Preview reached OpenRouter on Oct 8 and became instantly routable: a 600B-param sparse MoE (27B active) with 1M context, text+image+video input, at $1/$2.70 per 1M with cache hits at 10% of input โ an effective blended rate around $0.54/1M. AA Intelligence Index: 44 (#12) at ~$1.03 per task, just above Kimi K3 (43.6) and GLM-5.3 (44.8 is close) on the same board revision, with vendor-claimed DeepSWE 67.7%, GPQA 93.5%, FrontierFinance 66.4% (vs GPT-6 Astra's 55.0) and a Terminal-Bench v2.1 of 85%. Two honest catches: it is extremely verbose (160M output tokens across AA's run, ~2x median โ the real bill lives in output), and the aggregate trails the highlight charts (AA had it #12 while StepFun's cards cherry-pick finance wins). The unusual part is the open-weights date sitting in black and white: Oct 15 โ a date that came and went with no vendor repo on Hugging Face (the BF16 copies online are third-party mirrors) โ alongside the 24-hour H100 demo where the model optimized an MLA GPU kernel to 508 TFLOPS unsupervised, beating Claude Opus 5's 493. Open-weight frontier models just gained a credible Western-route competitor to Xiaomi's MiMo-V2.6-Pro at less than half the price.
Mistral Large 4 โ the trillion-param 'Le Chonk'
Mistral launched Large 4 (Oct 6, public preview) as its largest model ever: 1.05T-param MoE, 52B active per Mistral's current model docs (launch posts said 49B), 1.6B vision encoder, natively multimodal, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacentres โ the first model out of its โฌ3B round. Preview API: $0.68/$2.09 per 1M ($0.07 cached) against struck-through list pricing of $1.36/$4.18 with no announced end date. AA Intelligence Index v4.3.2 preview: 38.4. The pitch is sovereignty plus refusals: 82% on AA's find-and-patch cyber test โ the highest of any model, in a brief that also ranks ML4 top-five globally on AA's Cyber Index and leading open-weight models developed outside China 'by a wide margin' โ plus 93% on Cybench, SOTA open-weight SciCode-Verified, and visual grounding above GPT-6 Astra (Dense 200: 42% vs 41%). DeepSWE 61.7%, Coding Agent Index 49.8%, second of five in a blind Surge AI human eval behind Opus 5. The RL run continued past preview โ and delivered: weights went live Oct 29 (Mistral-Large-4-1T-A52B, 1.06TB bf16 on Hugging Face, no access gate), with three weeks of extra RL moving it to the open-weight coding lead (FrontierCode 2.0 19.7%, Terminal-Bench 4.0 41% โ +13 points, AA Coding Agent Index 48%, tied with GLM-5.3 at 47% fewer tokens per task) and 51% on the AA Cyber Index with a 96.6% refusal rate on malicious cyber requests โ Mistral's stated highest among open models. Independent AA Intelligence Index v5.0 rebased it to 25.3; pricing unchanged at $1.36/$4.18.
Liquid d1 โ a model that never generates tokens
Liquid AI's d1 (Oct 5) isn't an LLM and doesn't compete on the IQ leaderboards: it reads text/images in a single forward pass and outputs probabilities (yes/no, choice, score) in 200-300ms with zero generated tokens. Priced at $0.04 per 1M input tokens only โ and on six practical apps (support-ticket filtering, code search, filing 105 documents, web-agent action selection, context compaction, circuit-board defect inspection at 85-97% accuracy) it matched or beat GPT-6.1 Sol on four, at 19-200x lower cost. Images bill as input (~1,536 tokens for 1024x1024). Live on Liquid API, OpenRouter, Vercel. A separate API category โ decisions, not prose โ is now a product line.
Kolibri โ sovereign AI made in Germany
Aleph Alpha's Kolibri (Oct 3) is a different open-weight bet: not leaderboard-chasing, but EU-compliance-first deployment customer-controlled infrastructure. 78B-param MoE (3.46B active), 1M-token context validated (262K native, SWA position extension), reasoning mode + tool calling + RAG that abstains when evidence is missing. Pretrained on 20T tokens (~24% German) on 768 B200s โ 392k GPU-hours, 6.4e23 FLOPs โ with a German-optimized tokenizer. Apache 2.0, runs at ~78GB FP8 on 2x A100. Trained-data governance documented: 4.5M-URL blocklist (EC Piracy Watch List source), per-dataset license and opt-out screening. #1 on HF trending (600+ likes); Aleph Alpha is also merging with Cohere pending regulatory approval.
DeepSeek V4.1-Flash-1017 โ the checkpoint drop that ate a Friday
DeepSeek's Oct 17 release is the shape of an incumbent's update: no new model name, no new price, no fanfare โ a checkpoint swap that went to 558 points on Hacker News anyway. DeepSeek-V4.1-Flash-1017 (the date is the name) posts over the September Flash checkpoint in place: same 552B-param Causal Encoder-Decoder MoE (8B active prefill, 16B decode), same 1M context and native vision, same ~510GB FP8 MIT weights on Hugging Face, and the deepseek-flash API endpoint now serves it at the unchanged rates โ $0.15/$0.60 off-peak, $0.30/$1.20 peak, cache hits $0.003/$0.006. The gains are agent-shaped and vendor-reported: DeepSWE v1.1 74.2 โ 77.0 (the card's comparison table puts it above Claude Opus 5's 74.0), Terminal-Bench 4.0 31.2 โ 42.3, Automation-Bench 54.8 โ 60.5, HLE w/o tools 36.8 โ 42.6, with shorter outputs and faster task completion โ 'advancing the performance-cost Pareto frontier.' Independent AA intelligence numbers still cover only the September checkpoint (39.5 max), so treat the new table as DeepSeek's own until a refresh. The quieter story alongside: DeepSeek opened its Search and Fetch web-retrieval APIs to all users at $0.75 per 1,000 successful requests (5 QPS cap), with server-side calls through DeepSeek Harness, Claude Code or Codex free until Nov 1 โ an agent's browser, priced like bulk email.
Reflection Beam โ the US open-weight counter-attack
Reflection AI released Beam (Oct 5), its first open-weight model โ and made good a week early: the weights went live Oct 16 on Hugging Face (reflection/Beam-1.0-501B-A23B) under Apache 2.0, with instruction-tuned BF16, FP8 and NVFP4 shards, day-one vLLM/Transformers support and a dflash2 drafter for speculative decoding. 501B-total MoE, 23B active, pretrained on 23.8T tokens, then an RL campaign of 100M+ rollouts on 10.5K GB300 GPUs โ one of the largest open-lab RL runs ever, with 110K concurrent rollouts and ~1.3B sandboxes โ trained to penalize token waste: ~80% on Terminal-Bench 2.1 at ~25K reasoning tokens per attempt, where GLM-5.2 spends 80K for similar scores. Vendor card: SWE-Bench Verified 80.9, SWE-Bench Pro 65.5, DeepSWE 44.4; Artificial Analysis runs it at 30 on the Intelligence Index โ #2 among US open weights, behind Meta Muse Spark Open's ~39 but ahead of NVIDIA's Nemotron Super VL (26). The pricing chapter resolved within days: OpenRouter now lists Beam at $1.00/$3.00 per 1M (cache $0.10) with cheaper partner routes โ Nebius $0.25/$1.00, Together $0.40/$1.20 (cache $0.04), Baseten $1.30/$4.30 โ plus Reflection's own first-party API in public beta, free to start. At $1/$3 list it's a third of Sonnet-class money for a 501B open coder; a full self-host still takes the multi-GPU rack (H100/H200 nodes at $20-35/hour per third-party cost guides). Built by ex-DeepMind founding team (Mislav Balunoviฤ, Wei Chen โ AlphaZero/AlphaTensor lineage) with a $25B valuation, SpaceX Colossus-2 compute and a 250MW Korean sovereign AI factory. Forecasters had it for 2027 โ it shipped three months early.
Sonnet 5.5 โ the mid-tier ate the flagship's homework
Claude Sonnet 5.5 (Sep 28) keeps Sonnet 5's pricing ($2/$10) yet scores 56 on the AA Intelligence Index โ #2 overall, behind only Opus 5.5's 58 and at a quarter of Fable 5.1's $10/$50 price. Terminal-Bench 4.0: 70.6% at max effort vs Sonnet 5's 10.3% โ and above Opus 5.5's 66.4%, on a benchmark where its own flagship loses. It runs ~134 tok/s (its fastest Sonnet) and costs up to 30% less per task via fewer tokens; the catch is verbosity โ AA recorded the highest output tokens per task it has ever measured. Oct 8: cache reads halved to $0.10/1M, cutting most agentic workloads another ~20%. First Sonnet to ship with Opus-class cyber safeguards. Fable 5.5 is reportedly already in internal testing.
GPT-6.1 Sol โ the quiet DevDay upgrade, and the Astra that wasn't
DevDay (Sep 29) shipped GPT-6.1 Sol: same $2/$10 as GPT-6 Sol, cached input halved to $0.10, and AA Index 52 vs 48 at the same price โ within 3.5 points of GPT-6 Astra on 8 of 10 charts at 8โ23% of Astra's cost per task, matching Astra's best DeepSWE result (75.2%). Terminal-Bench 4.0 jumps 43.9% โ 56.1% for free. The bigger story was what didn't ship: OpenAI cancelled the GPT-6.1 Astra upgrade over safety regressions hours before its own conference. The Ultrafast tier is live for Astra (up to 6x faster in the API, 8x in Codex); 6.1 Sol Ultrafast is 'coming soon'.
Gemini 4 Argon โ Fairwind to GA in 13 days
Google announced Gemini 4 Argon on Sep 30 with intro pricing of $2/$10 and a 1M-token output limit (up from 64K) โ then restricted the whole launch to Fairwind Program cyber defenders and US-government pre-release testing. On Oct 13 it went GA: Gemini API paid tier, AI Studio, Antigravity (for Ultra subscribers) and the Gemini app/AI Mode in Search (Pro/Ultra). Intro pricing $2/$10 with cached input at $0.10 runs through Nov 30, 2026, then doubles to $4/$20 from Dec 1 โ landing it next to Opus 5.5 and Muse Wonder. The new spec is an industry-first 1M unified input-output token limit via a continuation protocol. Where it wins: #1 on the Vals Index (68.9%), Finance Agent v2 65.4%, DeepSWE v1.1 77.9%, Harvey Legal 19.6% (Astra: 5.4%), and ERQA 87% โ the frontier's best physical/spatial reasoning score, worth its own robotics launch piece. Where it trails: Terminal-Bench 4.0 57.4% vs Opus 5.5's 66.4%, and AA's Index puts it at 53 โ tied with GPT-6 Astra and Fable 5.1, below Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56). Google's own examples are unusually concrete: a Chromium reorganization spanning ~6M lines of feature code, ADK performance work folded into the public repo, 134 stale Chromium bugs triaged in under three hours, and Gemma 4 E2B serving up from ~37,900 to 108,000 tok/s on an A100. Successors 'Barium' and 'Carbon' are reportedly already in internal testing โ the intro price may expire before its successor ships.
Opus 5.5 โ a new #1 and the new value frontier
Claude Opus 5.5 (Sep 22) took the AA Intelligence Index top spot at 58 max on v4.3.2 โ now bracketed within its own family by Sonnet 5.5 (56 at $2/$10, Sep 28) โ while cutting price to $4/$20 (20% under Opus 5) and cache reads 60% to $0.20. Output is 30%+ faster, Fast mode runs $8/$40 at 2.5x, and Anthropic's tests put total run cost 40% below Opus 5. It leads six of ten Index evals (AA-Briefcase 1822 Elo, GDPval-AA 1846, HLE 61.4%, SciCode 66.9%) and posted the best automated-alignment scores Anthropic has measured โ while shipping with Fable-5.1-class safeguards because its bio/cyber capability now matches Mythos 5.1. Sonnet 5.5 wins Terminal-Bench 4.0 (70.6% vs 66.4%); Opus 5.5 stays the pick for complex open-ended work. Haiku 5.5 arrives in the coming weeks.
GPT-6 Sol & Luna โ OpenAI folds its own price card
One hour after Opus 5.5 landed on Sep 22, OpenAI launched GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50) โ 50% below the GPT-5.6 promo rates, with cached input at $0.20/$0.01 and batch at half list. Both are trained with Astra's methods; OpenAI claims Sol beats Opus 5 at ~9% of the cost, and GPT-6 Sol (AA 48) already outscores GPT-5.6 Sol (47) at half the price. Sep 29: GPT-6.1 Sol superseded GPT-6 Sol at the same price with cache at $0.10 and AA 52 โ routing built on GPT-6 Sol should already have been switched. GPT-5.6 Sol's $4/$20 promo now runs at least through Nov 21 with cheaper successors sitting below it.
MiMo-V2.6 โ open weights take the intelligence crown
Xiaomi released and open-sourced MiMo-V2.6 on Sep 22: Pro is a 1.02T-param MoE (42B active) scoring 46.32 on AA v4.3.2 โ the strongest open-weight model measured, ahead of GLM-5.3, Kimi K3 and Grok 4.7's 46 โ at $0.435/$0.87 with $0.0036 cache hits. Flash (310B/15B active) costs $0.14/$0.28. Native omni-modal input (text/image/video/audio), 1M context, MIT weights with the tech report and RL training resources published. Xiaomi broadcast parts of the RL run and claims 1/20โ1/60 of overseas pricing at equal intelligence.
Grok 4.7 โ same price, double the bill
Grok 4.7 (Sep 21) finally shipped after four missed windows at the same $2/$6 as Grok 4.6. It gained +2 to AA 46 and +111 Elo on AA-Briefcase (1657, just behind Opus 5 and Fable 5.1), with frontier price-performance on CursorBench 4.0. But AA measured 81k output tokens per task at xhigh โ double Grok 4.6's 38k, versus 27k for GPT-6 Astra max โ so identical per-token pricing can still mean a higher bill per finished task.
Qwen3.8-Omni-Flash โ the omni-modal price floor
Alibaba's Qwen team shipped Qwen3.8-Omni-Flash on Sep 18 โ its first model trained specifically for tool use in multimodal contexts. Text, images, audio (up to 3 hours) and video (2-hour files) go in as native inputs; text comes out. QwenCloud prices it at $0.15/$0.47 per 1M with $0.016 cache hits โ roughly a fifth of Gemini 3.8 Flash's intro rate, which itself doubles on Jan 1, 2027. Qwen's own tables put UniClawBench at 69.6 vs Gemini's 69.0, while Gemini keeps the lead on AgenticVBench (45.0 vs 36.8). Audio input prices fell 98% vs the previous omni tier. The catch: it's API-only โ the open-weight playbook stopped at the Flash-Next base.
Mercury 2.5 โ diffusion LLM at 1,107 tokens/sec
Inception Labs released Mercury 2.5 on Sep 8 โ the largest diffusion language model trained to date. It generates tokens in parallel rather than left-to-right, hitting 1,107 tok/s on widely-available NVIDIA GPUs. Quality is up 40% over Mercury 2, comparable to GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5. List $0.20/$0.75 per 1M with an 80% launch discount ($0.04/$0.15) expiring early October. 260K context, tunable reasoning, parallel tool calls, schema-aligned JSON. A diffusion LLM is now beating token-by-token models on the latency axis.
Cognition SWE-2 โ frontier coding without an API
Cognition shipped SWE-2 on Sep 10, post-trained from Kimi K3 with RL scaled to the multi-trillion-parameter regime. 50.0% on FrontierCode 1.1 Main โ within a point of Fable 5.1 (50.9%) at a claimed 64% lower cost, and 92.8% on Terminal-Bench 2.1, the best measured score. Free on Devin paid tiers through early October, then $3/$15 list pricing. No public API or model card โ SWE-2 is a product model, not a commodity. The open frontier keeps closing: the base model alone was already #7 overall.
GPT-6 Astra โ OpenAI's $10/$50 flagship
OpenAI announced GPT-6 Astra on Sep 3 and shipped the API model Sep 4. $10/$50 per 1M โ 2.5x GPT-5.6 Sol's $4/$20 promo โ with $1 cached input, $12.50 cache writes, and long-context rates above 272K tokens (2x input/cache, 1.5x output). Batch/Flex at 50%, Fast mode at 2x. 1.05M context, 128K output. Saturates ARC-AGI-3 (99.9%) and FrontierMath Tier 4 (98%); ExploitBench 100% without safeguards. First model to hit OpenAI's critical cybersecurity threshold. AA Intelligence Index: 61, but ~70% more token-efficient than Sol โ less than half Fable 5's cost per coding task.
Fable 5.1 โ the cache-read cut is the story
Claude Fable 5.1 (Sep 1) keeps $10/$50 but cuts cache reads 75% to $0.25/1M. Anthropic's own usage data: ~25% total savings on typical workloads, up to ~45% for agentic work where cache hits dominate. Terminal-Bench-Science 52.6% (more than doubles Fable 5), AutomationBench 31.4% (vs 17.1%), AA Index 66 โ #1. Candid caveats from the system card: forced tool use now returns 400, thinking blocks are model-bound, and Mythos 5.1 is 'less honest under pressure' than recent Claude models.
Gemini 3.8 Flash โ fourth Flash in four months
Google shipped Gemini 3.8 Flash on Sep 2 โ the fourth Flash since May (3.5, 3.6, 3.7, 3.8). AA Index 59 (+3 over 3.7) at the same $0.75/$3.75 intro through Dec 31, then $1.50/$7.50. ~300 tok/s โ the fastest output speed Artificial Analysis has measured. Trained on long-running agentic loops, it 'works harder': ~40% higher cost per task than 3.7 (30% more output tokens), so use low effort for efficiency-first routing. Gemini 3.8 Flash Cyber reaches trusted defenders via the Fairwind Program โ 2.6x more correct Chrome patches than larger frontier models.
The price floor keeps collapsing
Three weeks after four frontier launches (Sep 1โ3), the commercial floor dropped again: DeepSeek V4.1-Flash (Sep 10) at $0.15/$0.60 off-peak with cache hits at $0.003/1M โ 50x below its own cache-miss rate โ and Inception's Mercury 2.5 diffusion LLM (Sep 8) at $0.04/$0.15 launch pricing running 1,107 tok/s. Cognition's SWE-2 (Sep 10) matches Fable 5.1 within a point on FrontierCode at ~64% lower cost. GLM-5.3-Flash returned to list ($0.15/$0.50) on Sep 10. Calendar: GPT-5.6 Sol promo ends Nov 21 ($4/$20 โ $5/$30), Gemini 3.8/3.7 intro ends Dec 31 (doubles Jan 1, 2027), Mercury 2.5's 80% discount ends early Oct.