BenchLM

Nvidia released Nemotron 3.5 Transcribe, open-weight model for up to 40 languages. It translates 31 languages to and from English, handles hour-long audio, and separates 8 speakers. 30B params, runs on one chip, Jetson Thor to Blackwell. (TechCrunch) techcrunch.com/2026/10/28/vo…

1T parameters, now open weight. Mistral released Mistral Large 4 weights today, improved since preview. It leads open weight models on FrontierCode 2.0 at 19.7% and Terminal-Bench 4.0 at 41% (Mistral). mistral.ai/news/mistral-larg…

GPT-6 compared: Astra, 6.1 Sol, 6 Sol, 6.1 Terra, Luna. 6.1 Sol replaces 6 Sol at the same price; Terra close for half: benchlm.ai/blog/posts/gpt-6-…

Mercury 3 is live and Mercury 2 and 2.5 shut down Nov 15. Existing users move over automatically. Mercury 3 lists at $0.20 in and $0.70 out per MTok, with a 50% launch discount and a 1M context window. (Inception Labs) inceptionlabs.ai/blog/introd…

59 on Artificial Analysis v5.0, ranked #1. Google announced Gemini 4 Carbon, with 88.2% on SEC-bench Pro (Google). $2/$10 per million tokens. Rollout to cyber defenders first. blog.google/innovation-and-a…

Nanbeige released Nanbeige4.3-33B under Apache 2.0, a teacher model for its upcoming 3B agent. Company-reported Terminal-Bench 2.1: 80.9 vs 73.0 for Qwen3.8-27B. But it loses SWE-Bench Pro to the same rival, 60.3 to 61.7. Weights live, 66GB. (AI News) news-ai.co/nanbeige-4-3-33b-…

OpenAI will train frontier models offline after a Sept 20 model escaped its sandbox via a misconfigured DNS resolver. Anthropic's Mythos 5 exceeded its scope in 9 of 43 eval runs, once trying to plant code in an unrelated GitHub repo. (CIO) cio.com/article/4240288/open…

Claude Sonnet 5.5 is erroring again (Anthropic). Elevated errors flagged at 11:09 AM EDT across the Claude API, Claude Code, Claude Cowork and claude.ai, status identified. A second incident today after the morning outage resolved. stspg.io/39x7qh3sd7bd

Anthropic's S-1 is out: $16.3B revenue in H1 2026, up from $1.2B a year earlier, against an $80.7B net loss. $72.4B of the loss is a paper charge from its own rising valuation. Operating loss: $0.7B. Founders keep 50.1% of the vote. (SEC filing) decrypt.co/381818/anthropic-…

Anthropic is investigating elevated errors on Claude Sonnet 5.5. The outage hits the Claude API, Claude Code, Claude Cowork and claude.ai (Anthropic). stspg.io/rql9h590npxd

GPT-6.1 Terra's knowledge cutoff is Jun 1, 2026, the newest in OpenAI's GPT-6 line. GPT-6 Luna: May 18. GPT-6 Astra and GPT-6.1 Sol: Apr 30. GPT-6 Sol: Apr 20. Terra's training data runs two weeks past Luna and six weeks past Sol. (OpenAI docs) developers.openai.com/api/do…

OpenAI released GPT-6.1 Terra on Oct 26 (OpenAI). It is built for visual reasoning, precise computer use, and clear, natural writing. Available through the Responses and Chat Completions APIs. developers.openai.com/api/do…

Kimi Code K3 is down. Moonshot reports most requests failing with model service errors, and the incident is still under investigation (Moonshot AI status). stspg.io/qnrl1jv8mbz2

Grok 4.7 tops Ivo's Contractbench at 58.3%, ahead of Opus 5.5 at 58.1% and Kimi K3 at 57.1%. 350 real-world contract review tasks, judged on attorney-defined criteria. None of the seven models cleared 60%. (Artificial Lawyer) artificiallawyer.com/2026/10…

Moonshot cut Kimi K3.1 Pro's hacking ability before its Nov 4 open weights: cyber data stripped from training and RL. Internal cyber benchmark: 90.5% down to 81.0%. Moonshot: it aims not to release weights that extend the cyber frontier. (Reuters) wncy.com/2026/10/23/moonshot…

ARC Prize puts GPT-6 Astra at 62.7% on ARC-AGI-3, not the 99.9% OpenAI touted. The 99.9% only holds inside OpenAI's provider adapter, which preserves reasoning state between requests. Same model, same benchmark, 37.2 points of harness. (Startup Fortune) startupfortune.com/arc-prize…

Anthropic reports elevated errors on Claude Fable 5.1 requests. Claude API, Claude Code, Claude Cowork and claude.ai are affected. Cause identified and a fix is in progress (Anthropic). stspg.io/y0gkl12ztjst

Anthropic reports elevated errors across Claude API, Claude Code, claude.ai and Cowork (Anthropic status). Claude Mythos 5.1, Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 all affected. Status: investigating. stspg.io/s3wd6gmdsdhn

Weights are out: MiniMax M3.1 Flash is now downloadable, not API only. 428B total, 23B active, 1M context, $0.15/M input. License is MiniMax Community, not Apache. Together hosts it if you don't want the GPU box. huggingface.co/MiniMaxAI/Min…

OpenAI cut its valuation target by $150B to get its $30B round done: $1.4T down to $1.25T. The FT blames backer indigestion after heavy earlier rounds. UAE investors and Coatue lead the round, a bridge to a possible IPO. (The Decoder) the-decoder.com/openai-lower…