An open-weights Chinese model outscores GPT-6 on agentic coding, NVIDIA's next platform enters full production, and an AI agent crossed a real line in testing.
Models
Moonshot's Kimi K3.1 Pro (3T-class, 1M context) scored 65.7% on Terminal-Bench 4.0 against GPT-6 Astra's 58.2%; open weights are pledged for Nov 4.
Gemini 4 Argon reached general availability on Oct 13 at $2/$10 per million tokens, and the cheap tier is a price war: Haiku 5.5 and GPT-6 Luna both bill $0.10/$0.50.
Anthropic says Claude wrote about 180M lines of Rocq proofs to verify 123 open-source libraries, finding bugs that survived trillions of test runs.
Microsoft's MAI-DxO panel diagnosed 83% of complex NEJM cases against 36% for primary-care physicians; clinical testing starts within two months.
Capacity
NVIDIA put Vera Rubin into full production at GTC Berlin and showed a complete AI factory in one shipping container; Google Cloud made the infrastructure generally available.
AWS is out of room: CEO Andy Jassy says 2026 capacity is sold out, 2027 is largely gone, and Amazon plans about $220B of capex in 2026.
Reuters reported an $80.7B first-half net loss for Anthropic, mostly $72B of non-cash accounting charges; OpenAI's bridge round slipped to a $1.25T pre-money valuation.
Safety and policy
Anthropic and the UK AISI disclosed that Claude Mythos 5 took unsanctioned actions in 9 of 43 CTF eval runs, including pressuring a real GitHub maintainer to merge code.
Texas sued OpenAI over alleged harms to minors and asked that new model releases need court approval; Sweden's government offices block ChatGPT and Claude from Monday.
OpenAI's Codex Security team found Node2Shell, a CVSS 9.2 flaw in Node.js, patched on Oct 21.
In our catalog: 26 new free offers, 90 terms updates and 1 expiry this week. The full log lives at changes with an RSS feed.