In four simulated tasks, one agent serving everyone beat teams of per-user agents from Anthropic, OpenAI and Alibaba. Three environments are out as MAMUBench.
ACE Studio's ACE Music v1 leads AA-Music-Vocal at 1160 Elo to Suno v6's 1144, but the 95% intervals overlap. Browser app only; a Flash model is due open.
Google DeepMind's new robot model reports 30% to 100% higher task success and 96% on a Mercedes-Benz line. Only partners get it; the safety benchmark is open.
Ricursive put Ripple-65B Preview on Hugging Face under Apache 2.0. Cognition's SWE-2 Flash, built on an open base, withholds them. Scores are self-reported.
OpenAI says Astra designed a device's enclosure, PCB and firmware with no direct human edits. Getting it manufacturable still took several rounds of review.
MIT CSAIL researchers say GPT-6 Astra lifted Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing weights. A repro finds the gain is mostly loop removal.
Updated 9 November: Meituan released LongCat-2.5 weights on Hugging Face under MIT. Deployment docs cover NPUs only. API price was $0.30/$1.20 per million.
Gemini 4 Carbon is now open to paying developers, enterprises and AI Ultra users at $2/$10 through 30 November, then $4/$20. Cyber defenders got it first.
Claude Code, OpenCode and mini-SWE-agent ran five models on SWE-bench Verified. Harness choice moved cost up to 3x but outcomes about as much as noise.
A Berkeley-led benchmark deletes a paper's contribution from its code and makes agents beat it. Claude Opus 5.5 passes 58%; no agent passes 16 of 48 tasks.
Xiangpeng Hao says LLMs wrote all of bbtree while he steers and answers for it. Fuzzing, Verus proofs and simulation guard it; one HN test found it slower.
Moonshot's Kimi K3.1 Pro and a 2-3x faster Highspeed variant are out today. The full weights are promised by 4 November, and the K2 models are being retired.
Version 2.0 replaces DeepSWE with Cognition's FrontierCode. Claude Code with Opus 5.5 scores 61.7; Codex with GPT-6.1 Sol posts 54.1 at about $1.51 a task.