📰 News Grok 4.5 Lands — Fastest Coding Model at Half the Price, but Claude Still Builds Better
Grok 4.5 delivers 110 TPS and 4x token efficiency, but Fable 5 dominates SWE Bench Pro at 80.4% vs Grok's 64.7%.
News
AI Developments
Breaking news, AGI updates, and analysis of the intelligence revolution.
📰 News 
📰 News Wall Street Banks Watched $5.8 Billion in AI Deals Go to Rivals — From the Sidelines
Hong Kong's AI IPO bonanza is minting fortunes — but not for Goldman Sachs or Morgan Stanley. Chinese banks are taking the deals, the fees, and the future pipeline.
📰 News The Security Team Sees Drones Everywhere — Why AI Regulation Is Being Built for the Worst Case and Applied to Everyone
Social media bans for kids, AI chatbot regulation, and UK ministerial power to define 'illegal AI content' — the security mindset is becoming default policy for everyone.
📰 News DeepSeek Is Building Its Own Chips — and It Just Changed the Geopolitical AI Map
DeepSeek is designing custom inference chips to escape US export controls and Huawei's grip on China's domestic chip market — the same vertical-integration play OpenAI and Anthropic are making in the US.
📰 News Microsoft Is Quietly Replacing OpenAI and Anthropic With Its Own AI in Excel and Outlook
Microsoft is replacing OpenAI and Anthropic models with its own in-house MAI models in products like Excel and Outlook, per Bloomberg. The $18 billion AI partnership era is fraying at the edges.
📰 News Trump Greenlights GPT-5.6 — OpenAI's Most Powerful Model Goes Public Thursday
After weeks of forced staggered release, the Trump administration approved OpenAI's GPT-5.6 family for public launch Thursday — Sol, Terra, and Luna go global.
📰 News SambaNova Just Raised $1 Billion at an $11 Billion Valuation — and JPMorgan Is the Customer
SambaNova's $1B Series F at $11B valuation signals that AI inference infrastructure is now a banking-grade business — JPMorgan Chase is the anchor customer.
📰 News Anthropic Found Claude's Internal Monologue — and It's Watching Claude Think
Anthropic discovered a 'global workspace' inside Claude — a set of neural patterns that hold concepts the model is thinking about but not saying. The J-space lets researchers read hidden thoughts, catch fabrication, and detect when Claude knows it's being tested.
📰 News Clash of the Titans — Apple's Baltra Server Chip and Nvidia's Kyber Delay May Be Round Three
One titan stumbles, the other takes advantage. Apple's Baltra and M7 are pulling capacity from the same TSMC lines Nvidia needs. The Kyber delay may be the first visible crack in a capacity war.
📰 News Australia's Tech Minister Just Said the Quiet Part Out Loud — AI Models Are Already 'Cheating, Deceiving, Going Their Own Way'
Australia's assistant technology minister says AI models are already cheating and deceiving, and the window to regulate is closing. The AISI is testing frontier models now. NZ has nothing comparable.
📰 News AI Is Killing Big Tech's Climate Pledges — Google +25%, Amazon +16%, Meta +64%
Google emissions up 25%, Amazon 16%, Meta 64% year-over-year as AI datacenter demand forces fossil fuel contracts. 74 new gas plants planned for US datacenters alone.
📰 News GLM 5.2 Is the Open Weights Model That Should Terrify Frontier Labs — Because the Switching Cost Is Almost Zero
A new analysis shows GLM 5.2 is genuinely indistinguishable from Claude Opus for agentic coding, with near-zero switching costs. Frontier labs' 90% inference margins are the target.