Articles and AI news from Siarhei Holub.https://sholub.devhttps://sholub.dev/news/2026-11-09-claude-fable-5-2-mythos-5-2https://sholub.dev/news/2026-11-09-claude-fable-5-2-mythos-5-2The release is real and priced at $8/$40 per million tokens, but every benchmark in the post is Anthropic's own and no independent index has scored Fable 5.2 yet.Mon, 09 Nov 2026 20:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-meta-muse-emberhttps://sholub.dev/news/2026-11-09-meta-muse-emberTechCrunch reports Meta's new low-cost Muse Ember at $0.10 per million input tokens, but no Meta page was found and Artificial Analysis does not list it yet.Mon, 09 Nov 2026 20:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-xai-grok-api-watermarkhttps://sholub.dev/news/2026-11-09-xai-grok-api-watermarkxAI's docs say text from seven Grok models is statistically watermarked from 7 November unless a team admin opts out; how reliable detection is has not been tested outside xAI.Mon, 09 Nov 2026 20:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-mistral-large-4-weightshttps://sholub.dev/news/2026-11-09-mistral-large-4-weightsThe 1T-parameter weights are real and licensed as a modified MIT, but Mistral's coding and cyber numbers are mostly its own and Artificial Analysis ranks the model 11th of 26 open-weight models.Mon, 09 Nov 2026 20:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-gpt-6-1-terrahttps://sholub.dev/news/2026-11-09-gpt-6-1-terraOpenAI's docs confirm the price and Artificial Analysis independently measured a Coding Agent Index of 48 at $0.40 per task, though the launch post itself still does not open.Mon, 09 Nov 2026 20:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-gpt-6-cyberhttps://sholub.dev/news/2026-11-09-gpt-6-cyberOpenAI's post confirms the 48.5% ExploitGym score and 99.93% scope adherence. Irregular's independent run finds the model about on par with GPT-6 Astra on FrontierCyber but does not test ExploitGym.Mon, 09 Nov 2026 14:40:00 GMTnewshttps://sholub.dev/news/2026-11-09-openai-linux-kernel-exploitshttps://sholub.dev/news/2026-11-09-openai-linux-kernel-exploitsOpenAI's post counts 948 distinct local privilege-escalation bugs with working exploits, at about $1.2M. The figures are OpenAI's own, the patches are unreviewed candidates, and no outsider has verified them.Mon, 09 Nov 2026 14:40:00 GMTnewshttps://sholub.dev/news/2026-11-09-google-tpu-8tx-8i-maxhttps://sholub.dev/news/2026-11-09-google-tpu-8tx-8i-maxGoogle announced two more TPU systems with latency and bandwidth claims. All figures are Google's, several have no named baseline, and no availability date is given.Mon, 09 Nov 2026 14:30:00 GMTnewshttps://sholub.dev/news/2026-11-09-openai-acquire-pierrehttps://sholub.dev/news/2026-11-09-openai-acquire-pierreOpenAI announced it will acquire Pierre Computer Company to extend Codex's code review and Git tooling. Price and timeline are undisclosed, and the full announcement page could not be read.Mon, 09 Nov 2026 14:25:00 GMTnewshttps://sholub.dev/news/2026-11-09-llama-modchiphttps://sholub.dev/news/2026-11-09-llama-modchipAn open-source llama.cpp fork runs the 284B model in about 57 GiB by quantizing to 1.71 bits per weight. Speeds and quality are the author's own, and the format needs the project's own binary.Mon, 09 Nov 2026 14:20:00 GMTnewshttps://sholub.dev/news/2026-11-09-goldman-claude-code-restrictionhttps://sholub.dev/news/2026-11-09-goldman-claude-code-restrictionBloomberg-sourced reports say Goldman restricted AI coding tools for most traders and bankers in October. The original article was not read, the sources are anonymous, and Goldman's statement does not confirm the details.Mon, 09 Nov 2026 14:15:00 GMTnewshttps://sholub.dev/news/2026-11-09-cwe-bench-v2https://sholub.dev/news/2026-11-09-cwe-bench-v2A public 100-task benchmark where agents must close security flaws without breaking behavior. The leaderboard numbers come from the benchmark's own authors, and only three of eleven models are named.Mon, 09 Nov 2026 14:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-inkling-2-smallhttps://sholub.dev/news/2026-11-09-inkling-2-smallThe release and Artificial Analysis score of 31 are confirmed. The coding table is Thinking Machines' own, and the open-weights claim depends on a release that has not happened.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-deepseek-harness-usershttps://sholub.dev/news/2026-11-09-deepseek-harness-usersSelf-disclosed figures with an undefined metric and no outside measurement.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-swe-2-flashhttps://sholub.dev/news/2026-11-09-swe-2-flashThe release is real and in Devin, but every number is Cognition's own, the harness is not named, and the speed claim comes from a simulated replay.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-weight-space-programminghttps://sholub.dev/news/2026-11-09-weight-space-programmingA real, openly logged experiment, but author-run, unreviewed, and the big gain comes from removing reasoning loops in one tiny model.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-gemini-4-carbonhttps://sholub.dev/news/2026-11-09-gemini-4-carbonThe release is real and its #1 index rank is independently confirmed, but access is gated and the cyber scores are self-reported.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-co-design-bench-completion-gaphttps://sholub.dev/news/2026-11-09-co-design-bench-completion-gapThe strongest of ten tested models, GPT-6 Astra, passes 69.9% of acceptance criteria per task but fully satisfies only 26.4% of tasks.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-kimi-k3-1-pro-weightshttps://sholub.dev/news/2026-11-09-kimi-k3-1-pro-weightsWeights are on Hugging Face, but coding numbers are Moonshot's own, and the launch-day rank of fourth was later cut to ninth when Artificial Analysis changed its index.Mon, 09 Nov 2026 05:10:00 GMTnewshttps://sholub.dev/news/2026-11-09-frontiercode-leaderboardhttps://sholub.dev/news/2026-11-09-frontiercode-leaderboardOn real open-source tasks graded by whether a maintainer would merge, no model scores above 52% and pass rates top out at 58%.Mon, 09 Nov 2026 01:55:00 GMTnews