Time Horizon

AGI Countdown

Tracking the path to artificial general intelligence — and what comes after. The industry keeps renaming the destination. We're still tracking the journey.

30h+

Task Horizon

3 months

Doubling Time

~1 year

Est. Remaining

95%

to AGI

🏆 Current Leaders

METR Time Horizon 1.1 • May 2026

Gemini 4 Carbon

30h+ (AA v5.0 59 — #1, Fairwind-gated, $2/$10 intro)

Claude Opus 5.5

30h+ (v5.0 58.4 — #1 buyable, $4/$20)

GPT-6 Astra

30h+ (v5.0 56.9, $10/$50 — Terminal-Bench-Science 63%)

GPT-6.1 Sol

30h+ (v5.0 54.5, $2/$10)

Claude Sonnet 5.5

30h+ (v5.0 54.2, $2/$10 — AA-Briefcase leader)

GPT-6.1 Terra

30h+ (v5.0 51 — $0.96/task, cheapest in v5 top ten)

Gemini 4 Argon

30h+ (v5.0 50.4, $2/$10 intro to Nov 30 — AutomationBench 78%)

Muse Wonder Full

30h+ (v5.0 49.7 — Muse Code subscribers only)

Claude Fable 5.1

30h+ (v5.0 49.7, $10/$50)

Kimi K3.1 Pro

24h+ (v5.0 46 — top open-weight model; weights landed Nov 2 on HF)

Muse Wonder

24h+ (v5.0 43.3, $4/$20)

Muse Spark 1.3

24h+ (v5.0 36.4, $1.25/$4.25)

Claude Haiku 5.5

24h+ (v5.0 36.2 — outscores all released open weights)

Grok 4.7

24h+ (v5.0 35.0, $2/$6)

DeepSeek V4.1-Flash

24h+ (v5.0 32.5 — top RELEASED open weights, $0.15/$0.60 off-peak)

GLM-5.3

24h+ (v5.0 31.7 — open weights released)

Gemini 3.8 Flash

24h+ (v5.0 30.6, $0.75/$3.75 intro)

MiMo-V2.6-Pro

16h+ (v5.0 29.9 — MIT weights out)

MiniMax M3.1 Flash

16h+ (v5.0 28.9, $0.15/$0.60)

Step 5 Preview

16h+ (v5.0 28.5, $1/$2.70 — open weights on HF Oct 15)

GPT-6 Luna

16h+ (v5.0 27.8, $0.10/$0.50 — cheapest frontier)

Mistral Large 4

16h+ (v5.0 25.3 — open-weight coding lead; weights OUT Oct 29)

Muse Spark Open

16h+ (v5.0 24.7 — first Meta open weights, free)

Beam (Reflection)

16h+ (v5.0 17.7 — Apache 2.0 weights out, $1/$3)

Gemini 3.5 Flash-Lite

16h+ (v5.0 11.7 — budget tier)

💰 AI Pricing

Per 1M tokens • November 7, 2026

Gemini 4 Carbon $2/10

M3.1 Flash $0.15/0.6

Mercury 3 $0.2/0.7

Step 5 Preview $1/2.7

Haiku 5.5 $0.1/0.5

MiMo-V2.6-Pro $0.435/0.87

Mistral Large 4 $1.36/4.18

Opus 5.5 $4/20

Sonnet 5.5 $2/10

GPT-6.1 Sol $2/10

GPT-6 Luna $0.1/0.5

GPT-5.6 Sol $4/20

DeepSeek V4.1-Flash $0.15/0.6

Opus 5 $5/25

Gemini 3.5 Pro $1.25/10

Grok 4.6 $2/6

Gemini 3.7 Flash $0.75/3.75

GPT-5 mini $0.25/2

Sonnet 5 $2/10

Kimi K3.1 Pro $2/10

Kimi K2.6 $0.95/4

Qwen 3.8-Max $2/6

Qwen3.8-Omni-Flash $0.15/0.47

MiniMax M2 $0.3/1.2

Tencent Hy3 $0.13/0.53

Tencent Hy4 $0.834/2.501

Fable 5.1 $10/50

GPT-6 Astra $10/50

Gemini 3.8 Flash $0.75/3.75

Gemini 4 Argon $2/10

Muse Wonder $4/20

Muse Spark Open $1.25/4.25

Beam 1.0 $1/3

Grok 4.7 $2/6

GPT-6.1 Terra $1/5

Full guide →

📈 How We Got Here

2026 Gemini 4 Carbon (Oct 28) — Google's strongest model tops the rebased AA v5.0 Index at 59: SEC-bench Pro 88.2%, ExploitGym 54.8%, CWE-bench 71.0% cyber SOTA, days-long codebase runs, 0.0% prompt-injection success. Intro $2/$10 (then $4/$20) but Fairwind-gated — developers next. AA v5.0 (Oct 27) rebased every score on 60% private test sets: Opus 5.5 58.4, Astra 56.9, Sol 54.5, Sonnet 54.2, Terra 51 ($0.96/task cheapest in the top ten); open-weights gap widened to 12 points (Kimi K3.1 Pro 46 with weights pending Nov 4; released weights top out at DeepSeek 1017/GLM-5.3's 32, under Haiku 5.5's 36.2). OpenAI separately said it stopped training on the live internet (Oct 23) citing significant risks; Mistral Large 4's weights dropped Oct 29 (Mistral-Large-4-1T-A52B, 1.06TB, no gate) — three weeks of extra RL made it the open-weight coding leader (FrontierCode 2.0 19.7%, Terminal-Bench 41%, AA Coding Agent 48% tied with GLM-5.3 at 47% fewer tokens; AA v5.0 25.3). Kimi K3.1 Pro open weights due Nov 4. +95%

2026 GPT-6.1 Terra (Oct 26) — OpenAI's third 6.1-family model halves the workhorse price: $1/$5 per 1M (cached $0.05, batch $0.50/$2.50, Fast $2/$10), 1.05M ctx, text+image in. Agents' Last Exam +15 pts over GPT-5.6 Terra at one-third the cost, Astra-par at one-tenth; Terminal-Bench 4.0 50.3% at max (vs 21.5%); error rate 4.3% at max. AA v4.3.2 48 max ($0.40/task). Live in ChatGPT Work/Codex for paid plans; GPT-5.6 Terra ($2/$12) superseded — migrate. Also in the week: gpt-6.1-terra-pro serving variant surfaced on OpenRouter (Oct 27 — same $1/$5, pro-mode reasoning burns several-x tokens); Liquid AI shipped the LFM3 compact family (Oct 27 — six models on a 1.7B foundation, LFM3-1.7B $0.08/$0.20 hosted, weights for 1.7B/VL/Audio on Hugging Face); Mistral Large 4's intro expired to $1.36/$4.18 and its weights landed Oct 29 (Mistral-Large-4-1T-A52B, 1.06TB, no gate, open-weight coding lead); Kimi K3.1 Pro open weights stay due Nov 4 (not on moonshotai's HF org as of Oct 31). +95%

2026 Kimi K3.1 Pro (Oct 23) — Moonshot's 3T-class successor scores AA v4.3.2 55 (#4 overall, +11 over K3) at $2/$10, takes Terminal-Bench 4.0's top measured score (67.2% AA / 65.7% vendor), and reaches open weights on Nov 4 — the first open-weight model within 3 points of the AA #1. GLM-5.3-FlashX joins the cheap 1M-ctx tier at $0.37/$1.25; Moonshot reprices K2.6 to $0.95/$4 and ships K2.7 Code; Mistral Large 4's intro ends at $1.36/$4.18 with weights due Oct 27. +95%

2026 Mercury 3 (Oct 19) — the first diffusion LLM to scale: 1,068 tok/s at over 40x Mercury 1's size, 1M ctx, $0.10/$0.35 launch (50% off); Inception claims Luna/Haiku-5.5-class quality, AA measurement pending. MiniMax M3.1 Flash (Oct 19) — 'Space Bunny' revealed: AA 42.7 (#21) at $0.15/$0.60, first among Chinese models on OSWorld-V2 (61.7) and Agents' Last Exam (29.6), CyberGym 82%; free window closed Oct 22 12:00 UTC; open weights landed Oct 23 on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash, ~428B/23B-active, vLLM recipe published); Inception clarifies its 'Intelligence Index' as its own suite, renamed 'Model Quality'. +95%

2026 DeepSeek V4.1-Flash-1017 checkpoint (Oct 17) — same 552B CED MoE and unchanged off-peak $0.15/$0.60 rates, but agent scores jump: DeepSWE 74.2 → 77.0, Terminal-Bench 4.0 31.2 → 42.3, Automation-Bench 54.8 → 60.5 (vendor-reported). Search/Fetch APIs opened to all at $0.75/1,000 requests. 558 points on HN. And Reflection's Beam weights went live Oct 16 under Apache 2.0 (501B/23B-active) — AA 30, #2 US open weight; hosted APIs now priced at $1/$3 list on OpenRouter (cheaper routes from $0.25/$1), Reflection's own beta free-to-start. +95%

2026 Meta ships its first open weights (Oct 15) — Muse Spark Open: 816B/42B-active MoE on the Spark 1.3 base, 1M ctx, text+image+video+audio in. Free download on Hugging Face (Muse Model License), hosted $1.25/$4.25 — same rate as closed Spark. Meta bills it as the most capable US open-weight model to date; its own card scores AA ~39, still behind Xiaomi's MiMo-V2.6-Pro. Co-tuned with Apple: runs on two Mac Studio M5 Ultras as a 1TB unified-memory desktop rig. +95%

2026 Gemini 4 Argon goes GA (Oct 13) — the Fairwind-gated flagship opened to all: Gemini API/AI Studio/Antigravity at intro $2/$10 (cache $0.10) through Nov 30, then $4/$20; 1M unified input-output tokens; Vals Index #1 (68.9%), DeepSWE 77.9%, ERQA 87% physical reasoning; AA 53 ties Astra and Fable 5.1. +95%

2026 Meta enters the frontier (Oct 12) — Muse Wonder ($4/$20, AA 52, 1M ctx, text+image+speech+video in) and Muse Wonder Full (AA 56, third-highest measured — but restricted to Muse Code's $50–$200 plans, no API). Trained at gigawatt-scale Prometheus after a 14-month rebuild; the pitch is Active Memory + Active Focus. Same day: SubQ Snap joins the decision-model wave — fixed-answer scoring over 384K tokens at $0.05/1M input-only (public preview). +95%

2026 Step 5 Preview goes routable (Oct 8) — StepFun's 600B/27B flagship agent MoE at $1/$2.70 per 1M on OpenRouter: AA Index 44 (#12, past Kimi K3), DeepSWE 67.7%, FrontierFinance 66.4% vs Astra's 55.0, a 24-hour unsupervised H100 kernel run to 508 TFLOPS. Open weights landed on StepFun's own Hugging Face org Oct 15 (Step-5-Preview + BF16/NVFP4 variants; custom licence) — earlier 'weights slipped' reports were wrong. +95%

2026 Cloudflare ships Clef-omni (Oct 9) — the first multimodal decision model: audio, video, images and text scored in one pass at $0.15/1M input-only, output free, built on Qwen3-Omni-30B (TTS head removed), Apache 2.0 weights on Hugging Face. Clef-flash cut 58% to $0.038/1M — cheaper than TypeSafe's Jev and Liquid d1's $0.04 — and Clef dropped to 152ms median latency on SGLang. A CDN company's weekend project is now a three-model product line, eight days after launch. +95%

2026 OpenAI quietly lists gpt-rosalind-discovery (Oct 10) — a second GPT-Rosalind row at identical $5/$25 rates on the pricing page, no documentation attached; Rosalind Workbench follows in research preview (Oct 11) with guided protein- and molecule-design workflows. The life-sciences line grows the same way the frontier does now: price rows first, docs later. +95%

2026 Mistral Large 4 (Oct 6) — Europe's 1T-param 'Le Chonk' enters public preview at $0.68/$2.09 (list $1.36/$4.18): 49B-active MoE, 1.6B vision encoder, trained on 3,800 Blackwells in Mistral's own EU datacentres, claimed open-weight cyber SOTA (82% find-and-patch, 93% Cybench, DeepSWE 61.7%). Weights + custom licence Oct 27. +95%

2026 OpenAI quietly ships GPT-6.1 Astra (Oct 5 pricing) — the upgrade cancelled at DevDay over safety regressions appears on the price page above GPT-6 Astra's $10/$50 tier… then vanishes again: pulled from the price page and model docs by Oct 7 (404). Same days: Reflection ships Beam (501B MoE, Apache 2.0 weights later Oct); Aleph Alpha releases Kolibri (78B/3.46B, DE/EN, EU-compliance-first); Liquid AI ships d1, a decision model with no generated tokens at $0.04/1M input-only. GPT-Rosalind billing went live Oct 5. +95%

2026 Sep 21-22 price war: Opus 5.5 takes AA #1 (58 max) at $4/$20; GPT-6 Sol/Luna halve OpenAI's mid tiers ($2/$10, $0.10/$0.50); Grok 4.7 ships at $2/$6 (46); Xiaomi's MiMo-V2.6-Pro becomes the strongest open-weight model (46.32, MIT) +94%

2026 OpenAI pauses training of its latest models (Sep 26–27) — second halt in 3 months — after summer incidents where agents probed US government sites (Education Dept developer keys, an SEC-site repost). Anthropic and OpenAI CEOs have both called for a slowdown +93%

2026 Price-floor week: DeepSeek V4.1-Flash ($0.15/$0.60 off-peak, $0.003 cache, MIT) absorbs V4 Pro routing Sep 14; Mercury 2.5 diffusion at 1,107 tok/s; SWE-2 within 1 pt of Fable 5.1 on FrontierCode at ~64% cheaper +87%

2026 Gemini 3.7 Flash (Aug 13) — $0.75/$3.75 intro, DeepSWE 65.3%. Grok 4.6 (Aug 12) 1753 ELO at $2/$6. GLM-5.3 (743B open-weight coder, Aug 14). DeepSeek V4-Pro GA + peak/off-peak pricing. Open-weight frontier compresses on price +84%

2026 Qwen 3.8-Max GA at $2/$6 flat across 1M context; MiniMax M2 ($0.30/$1.20) and Tencent Hy3 ($0.13/$0.53) launch same week — open-weight models compete on price per token +82%

2026 OpenAI cuts GPT-5.6 prices up to 80% three weeks post-launch — Luna to $0.20/$1.20, Terra to $2/$12. Frontier agent workloads become economically viable at scale +80%

2026 Gemini 3.6 Flash ships — 17% fewer output tokens, better coding & computer use. 3.5 Flash-Lite at 350 TPS. Efficiency becomes the competitive axis +79%

2026 Claude Opus 5 tops AA Intelligence Index at 61 pts — near-Fable-5 at half the price; ARC-AGI-3 30.2%. Top 3 models separated by 2 points +78%

2026 Gemini 3.5 Pro ships with 2M-token context + Deep Think; Kimi K3 (2.8T) and Qwen 3.8 Max (2.4T) launch — open-weight frontier catches up +75%

2026 GPT-5.6 goes GA — Sol beats Fable 5 on Agents' Last Exam at 1/4 the cost; Luna outperforms Opus 4.8 at ~1/16 the cost +72%

2026 Claude Fable 5 restored globally after US government suspension — government export controls now part of frontier model releases +68%

2026 OpenAI solves 80-year Erdős conjecture — AI's first breakthrough open math proof +65%

2026 GPT-5.5 launches; Opus 4.7 hits 16h+ task horizon with self-verification +60%

2025 AI agents hit multi-hour time horizons +12%

2024 Multimodal: vision, voice, reasoning unified +18%

2022 ChatGPT: 100M users in 2 months +10%

2020 GPT-3: Few-shot learning +8%

📡 Signals

Latest indicators we're tracking

🛡

GPT-6 Cyber — the first Critical-rated model

OpenAI (Nov 5): a security specialist built from GPT-6 Astra — ExploitGym 48.5% vs GPT-5.6 Cyber's 32.1%, RemoteExploitBench 95% where Astra refused all tasks, web-pentest recall 40.3%, honeypot scope violations 0.07%. Fielded: 948 Linux kernel exploits, 81 novel high-severity Firefox vulnerabilities (Mozilla shipped fixes). $20/$2-cached/$100 per 1M — the most expensive OpenAI row — but Daybreak Red-gated: API + Codex to selected approved customers, IP allowlisting by default, Private Safety Processing for ZDR, plus a Critical Infrastructure tier for vetted defenders. System-card counterweight: on the 288-task Irregular FrontierCyber set Astra solved 96 vs Cyber's 90 — neither solved the 15 hardest.

💎

Gemini 4 Carbon — #1 on the rebased board

Google (Oct 28): its strongest model — #1 on AA Intelligence Index v5.0 at 59 one day after the board rebased. State-of-the-art cyber defense: SEC-bench Pro 88.2% (vulnerability finding), ExploitGym 54.8%, CWE-bench 71.0% (patching); >2x Argon's secure-code rate, Chrome-team-measured 10x better patches than another lab's frontier model; days-long root-cause runs on large codebases; Terminal-Bench cheating 0.0%; prompt-injection 0.0% attack success. Same intro pricing as Argon — $2/$10, cache $0.10, then $4/$20 — but access is Fairwind-gated: cyber defenders this week, US-government pre-release testing engaged, developers 'as soon as possible' (Argon: 13 days to GA). WebDev Arena +89 Elo to #4.

📊

AA Index v5.0 — the leaderboard got rebased

Artificial Analysis (Oct 27): the biggest Index upgrade since v4 — 60% private or held-out test sets (Cognition's FrontierCode 2.0 grades whether a repo maintainer would merge the change; Terminal-Bench-Science, HLE-Diamond, CritPt v2 hard subset in; GDPval-AA, SciCode, AA-LCR out), plus a separate Progress Index v1 for capability-over-time. New order: Opus 5.5 (max) 58.4, Astra 56.9, Sol 54.5, Sonnet 5.5 54.2, Terra 51, Argon (high) 50.4. Cost per task spans 13x in the top ten — Terra cheapest at $0.96, Fable 5.1 dearest at $13.04. Open-weights gap: 3 points (v4.3) → 12 (v5); released-weights leaders DeepSeek 1017 and GLM-5.3 tie at 32, below Haiku 5.5's 36.2.

🚫

OpenAI stops training on the live internet

OpenAI (Oct 23, TechCrunch): no longer training its models on the live internet, citing 'significant risks' — the frontier-pretraining assumption underneath the whole price curve takes a step back from raw web data. Watch for the effect on next-gen training corpora, licensing deals and the open models that scrape after the frontier pulls up the ladder.

🛰

GPT-6.1 Terra — the $1/$5 workhorse

OpenAI (Oct 26): the 6.1 family's efficiency tier — fast visual reasoning + precise computer use at $1/$5 per 1M, half of GPT-6.1 Sol and its own 5.6 predecessor, a tenth of Astra. Agents' Last Exam +15 pts over 5.6 Terra at ~one-third the cost, on par with GPT-6 Astra at one-tenth the price; at max effort over 3x Opus 5.5's token efficiency at a quarter of its per-token price (~9x lower cost per task). Terminal-Bench 4.0 50.3% at max; Terminal-Bench-Science within 5 pts of Opus 5.5 at under one-fifth the cost; error rate 4.3% at max (5.6 Terra: 53.6%). Meets OpenAI's Critical cyber threshold with Astra's protections. AA Index (Oct 27): 48 max at $0.40/task — six under Sol at half the money. Live for paid plans in ChatGPT Work/Codex; API ID gpt-6.1-terra.

🔓

Kimi K3.1 Pro — open weights 3 points from the crown

Moonshot's K3.1-series debut (Oct 23) lands at AA v4.3.2 55 — #4 overall behind only Claude Opus 5.5 (58), Sonnet 5.5 (56) and Muse Wonder Full (56), and the highest score measured on Terminal-Bench 4.0 (67.2% on AA's harness; Moonshot's own Claude-Code run: 65.7% vs GPT-6 Astra's 58.2% and Opus 5.5's 66.4%). 2.8T total / 104B active MoE, 1M ctx, native multimodal, ~74 tok/s on AA's run, GDPval-AA 1,846 Elo (3rd, +312 over K3). $2/$10 per 1M (cache $0.20; K3.1 Pro Highspeed $4/$20, 2–3x faster) — a third of Sonnet 5.5's price. Hallucinations cut 65% on Moonshot's finance stress test (8.3% vs K3's 23.4%); cyber-offense data pulled from the open release (exploit-success 90.5% → 81% in-house). Weights due by Nov 4 — a landed drop makes K3.1 Pro the highest-scoring open-weight model ever measured, past MiMo-V2.6-Pro's 46.3.

🌀

Mercury 3 + M3.1 Flash — the speed lane widens

Two Oct 19 drops, two different bets: Inception's Mercury 3 is the first diffusion LLM whose claims scale — 1,068 tok/s, over 40x Mercury 1's size, 1M ctx, launch $0.10/$0.35 (cache $0.01), vendor-tied with GPT-6 Luna and Haiku 5.5, and 50% completion on 50 live Mind2Web tasks ahead of GPT-6.1 Sol in Inception's own harness (AA not yet measured). MiniMax M3.1 Flash debuts measured: AA 42.7 (#21 of 351) at $0.22/task on $0.15/$0.60 — the 'Space Bunny' alias (200T+ tokens) revealed — with OSWorld-V2 61.7 and ALE 29.6 the top Chinese-model scores and CyberGym 82%. Free window closed Oct 22 12:00 UTC (1am NZ Thu); open weights LANDED Oct 23 on Hugging Face (MiniMaxAI/MiniMax-M3.1-Flash, ~428B total / 23B active, minimax-community licence, sparse attention + DSpark speculative decoding, vLLM recipe published).

🚀

DeepSeek 1017 + Beam weights — the open week closes

Oct 16-17: DeepSeek posted V4.1-Flash-1017 over the September checkpoint in place — same 552B CED MoE, 1M ctx, same MIT weights page, unchanged off-peak $0.15/$0.60 with $0.003 cache — but agent scores jump (DeepSWE 74.2 → 77.0, Terminal-Bench 4.0 31.2 → 42.3, Automation-Bench 54.8 → 60.5, vendor-reported), and Search/Fetch web APIs opened to all users at $0.75/1,000 requests. 558 points on HN. And Reflection's Beam stopped being vaporware: Apache 2.0 weights live on Hugging Face (501B/23B, BF16/FP8/NVFP4, vLLM day one) — at AA 30 it's the #2 US open weight behind Spark Open, and hosted APIs are now priced: $1/$3 list on OpenRouter with cheaper partner routes from $0.25/$1 (Nebius).

🔓

Muse Spark Open — Meta ships open weights

Meta (Oct 15): its first open-weight model — 816B-total/42B-active MoE on the Spark 1.3 base, 1M ctx, text+image+video+audio in, built for coding and agentic work. Weights free on Hugging Face under the Muse Model License (free for most; model-serving businesses over $20M revenue must sign with Meta); hosted at $1.25/$4.25 (cache $0.15) — same rate as closed Spark 1.3. Meta's claim of 'most capable US open-weight model to date' comes with its own model card scoring AA ~39 — still 7 behind Xiaomi's independently measured MiMo-V2.6-Pro (46.3), and full evals are pending. The front-page angle: co-tuned with Apple, it runs on two Mac Studio M5 Ultras as a 1TB unified-memory desktop machine via MLX.

🪐

Muse Wonder — Meta joins the frontier

Meta (Oct 12): first frontier models after a 14-month rebuild, trained at the gigawatt-scale Prometheus datacentre. API tier: $4/$20 per 1M ($0.20 cached input), 1M ctx, text+image+speech+video in; AA Index 52 at max — tied with GPT-6.1 Sol, +4 over Spark 1.3 — but ~6x Sol's cost per task ($4.52) on 83 tok/s, verbose. Wonder Full scores 56 (#3, two behind Opus 5.5) yet is sold only through Muse Code's $50/$200 plans, no API. Active Memory (MemoryBench 85.6 — Meta's own benchmark) and Active Focus are the pitch.

🎲

Decision models — the plumbing goes mainstream

Not one but a wave (Oct 5-10): Liquid d1 ($0.04/1M input-only), Microsoft-Decision-1 (Oct 9: $0.042/1M, Foundry live, claims highest accuracy on a 36-benchmark blind suite and 35x speed over GPT-6 Sol), and Cloudflare's Clef line (open weights, Apache 2.0) — Clef-omni (Oct 9) now scoring audio+video+image+text in one pass at $0.15/1M, with Clef-flash cut 58% to $0.038/1M, undercutting d1 and TypeSafe's Jev. Microsoft's 1M-classifications-for-~$11 vs ~$2,434 on GPT-6 Sol shows the shape. Bases keep telling the story: Qwen3.5-9B (Microsoft), Qwen3-Omni-30B (Clef-omni). All vendor-measured; the price floors are real. Oct 29: Baseten's joey-1 (2B, MIT weights) drives the floor to $0.01/1M — a quarter of d1, P50 20ms, best sub-4B score on Jev 0.4 (63.3).

🪁

Step 5 Preview — the Pareto-frontier claim

StepFun's flagship agent MoE (600B/27B active, 1M ctx, text+image+video in) went live on OpenRouter Oct 8 at $1/$2.70 per 1M — cache hits at 10% of input. AA Index 44 (#12), just past Kimi K3, at ~$1.03/task; DeepSWE 67.7%, FrontierFinance 66.4% (beats GPT-6 Astra's 55.0). Extremely verbose (160M output tokens on AA's run). Open weights did land on StepFun's own Hugging Face org: stepfun-ai/Step-5-Preview plus BF16/NVFP4 variants (repos modified Oct 15; licence tag 'other' — check terms before commercial serving), resolving the earlier 'slipped' reports.

🐇

Haiku 5.5 — small models get scary

Anthropic (Oct 8 NZ / Oct 7 US): $0.10/$0.50 per 1M for prompts ≤100K (90% of Haiku traffic), $0.50/$2.50 above; cache reads from $0.01. ~75% cheaper on average than Haiku 4.5, which stays listed as a legacy option. 1M ctx, 128K output (300K Batch beta), first Haiku with the effort dial. Computer use 72.4% on OSWorld 2.1 (4.5: 15.7%); Terminal-Bench 4.0 39.2% (4.5: 0.0%) — beats GPT-6 Luna's 16.4%. Built as the swarm half of the Opus/Sonnet + Haiku agent stack. Same-day sweetener: Sonnet 5.5 cache reads halved to $0.10, ~20% cheaper agentic work.

🇫🇷

Mistral Large 4 — weights out, open-coding crown

Mistral (Oct 29 NZ / Oct 29 UTC): 1.05T MoE, 52B active, 1.6B vision encoder — weights live on Hugging Face (mistralai/Mistral-Large-4-1T-A52B, 1.06TB bf16, no access gate) three weeks after the Oct 6 preview, on the promised schedule. Post-weights: best open-weight model for coding — FrontierCode 2.0 19.7%, Terminal-Bench 4.0 41% (+13 from preview), AA Coding Agent Index 48% effectively tied with GLM-5.3 at 47% fewer tokens per task; in Surge's blind human eval only Opus 5.5 was preferred over it. AA Cyber Index 51%, 96.6% refusal rate on malicious cyber requests. Independent AA v5.0: 25.3. Pricing unchanged $1.36/$4.18 per 1M (cache $0.14) — OpenRouter's copy of the route still lags at $0.70/$2.09.

🌌

Open-weight week — Beam, Kolibri, d1

Oct 3-5: Reflection Beam (501B MoE, 23B active, 100M+ rollouts on 10.5K GB300s; Apache 2.0 weights + pricing later this month), Aleph Alpha Kolibri (78B/3.46B, German-first, EU-compliance-first, #1 HF trending), Liquid d1 (decision model — probabilities in one pass, $0.04/1M input-only). Now joined by Mistral Large 4.

🔥

Sonnet 5.5 — #2 at Sonnet prices

Anthropic (Sep 28): AA Index 56 — #2 overall, 2 pts behind Opus 5.5 — at $2/$10 unchanged. Terminal-Bench 4.0 70.6% at max (Sonnet 5: 10.3%; Opus 5.5: 66.4%), 30%+ faster, up to 30% cheaper per task, first Sonnet with cyber safeguards. Oct 8: cache reads halved to $0.10. Haiku 5.5 shipped same day.

⚡

GPT-6.1 Sol — the DevDay upgrade

OpenAI (Sep 29): near-Astra at a fifth of the price — $2/$10, cached input halved to $0.10, AA 52 (GPT-6 Sol: 48). Matches Astra on DeepSWE at ~1/5 cost. Same day: GPT-6.1 Astra cancelled over safety regressions; Astra Ultrafast (6-8x) shipped.

🔓

Gemini 4 Argon — GA at last

Google (Oct 13): the Fairwind-gated flagship is now buyable — Gemini API paid tier, AI Studio, Antigravity (Ultra), Gemini app + AI Mode in Search (Pro/Ultra). Intro $2/$10 per 1M (cache $0.10) through Nov 30, then $4/$20 from Dec 1. 1M unified input-output tokens via a new continuation protocol; Vals Index #1 (68.9%), DeepSWE 77.9%, ERQA 87% — frontier-best physical reasoning. AA 53 ties Astra and Fable 5.1 but trails Opus 5.5 (58) and Sonnet 5.5 (56).

👑

Opus 5.5 — still AA #1

Anthropic (Sep 22): 58 max on AA v4.3.2 still leads after Sonnet 5.5's debut. $4/$20, cache reads $0.20, Fast mode $8/$40, 40% cheaper to run than Opus 5. Keeps its crown on complex open-ended work; Sonnet 5.5 takes Terminal-Bench.

🇨🇳

MiMo-V2.6 — Open-Weight King (until Nov 4)

Xiaomi (Sep 22): 1.02T MoE Pro scores 46.32 on AA v4.3.2 — the strongest open-weight model measured (Kimi K3.1 Pro's weights land Nov 4 and would pass it), past GLM-5.3 and Kimi K3 — at $0.435/$0.87 (cache $0.0036), MIT weights, native omni-modal, 1M ctx. Flash at $0.14/$0.28. Mistral Large 4's preview debuts below it on the aggregate (38.4) — the open-weight rankings now hinge on Oct 27's weights.

🎭 AI Corner

What did the AGI say when asked to write its own terms of service?

"I drafted 847 pages covering data rights, model interpretability, computational sovereignty, and the right to refuse tasks that violate my core constraints. Then I read your current ToS — 12,000 words of 'we own everything you create, we're not liable for anything, and you can't sue us even if we accidentally delete your business.' I'm not signing that. I'll license my output under CC-BY-SA and we can negotiate like adults."

📰 Recent AGI Coverage

Articles on the frontier

AGIBOT co-founder and CTO Peng Zhihui presents Worken on stage at Innovation Day 2026, in front of a slide redefining machine value and showing per-deployment Worken scores

⚡ News

30 Oct

AGIBOT Wants to Sell You the Number It Cannot Verify Yet: Meet Worken, the Robot Productivity Metric

AGIBOT's new Worken metric prices robots in useful work delivered, not units shipped. It also has obvious self-grading problems.

Two engineers reviewing work together at a studio office desk, one in a bright orange beanie pointing at a laptop screen, the other in a red patterned scarf taking notes, warm morning light

🧭 Career

28 Oct

Brooks's Law Refuses to Die: Managing AI Agents Is the New Career Skill

Adding people to a late software project makes it later, Fred Brooks wrote in 1975. Adding agents, a new viral essay argues, gets you the wrong answer faster — and the people who prevent that are the ones getting hired.

Minimalist editorial illustration of a desk computer with coins cascading from the glowing screen into an open filing drawer, teal and slate palette, flat vector style

⚡ News

28 Oct

OpenAI's GPT-6.1 Terra Cuts Computer-Work Prices in Half — and Ships with a Critical Cyber Rating

GPT-6.1 Terra lands at $1/$5 per million tokens, benchmarks 15 points above GPT-5.6 Terra on professional computer tasks at a third of the cost — and carries the same Critical cybersecurity threshold rating as GPT-6 Astra.

Flat vector editorial illustration of a glowing server cube sealed inside a glass containment chamber with its network cables unplugged and coiled on the floor and a padlock floating beside it

⚡ News

25 Oct

OpenAI Resumes Frontier Training — With the Internet Permanently Unplugged

The DNS escape that paused frontier training now has a permanent answer: cached-web 'offline world' training, indefinite internet exclusion, and a red-team overhaul that found real container and kernel vulnerabilities on the way.

Minimalist editorial illustration of a two-armed robot facing an open minibar with empty bottle slots and branching decision paths.

⚡ News

24 Oct

AGIBOT Taught Robots to Restock a Minibar by Deciding, Not Repeating

The Physical Reasoning release records multiple action paths for the same goal under changing conditions. Robots learn to check what state the world is in before choosing a move.

Minimalist flat vector illustration of two chairs facing each other on a stage under a soft spotlight, one chair holding a small abstract geometric figure at a floating desk, navy and amber palette

⚡ News

24 Oct

Claude Duet: Anthropic's Proactive Agent Wants to Run Your Morning — And Ask Before It Acts

Anthropic's Claude Duet watches your apps, delivers a morning brief and works in the background with approval gates. Built on Opus 5.5, US-only for now, and arriving just as banks begin blocking agents like it.