GPT-6 Astra is out, and the part that is not just another Critical cyber headline is the 3D and real-world jump. On BenchCAD it reconstructs CAD from multi-view renders at 95.9% geometric overlap against Sol’s 83.3% and Fable 5.1’s 84.3%, and on workplace-style automation it more than doubles Sol on AutomationBench at 41.4% versus 18.1%, while posting 92.7% on ScreenSpot-Pro and 72.6% on OSWorld 2.0. FrontierMath Tier 4 sits at 97.6% and Terminal-Bench Science at 64.6%, which is where the academic gap shows, but Artificial Analysis still has Fable 5.1 ahead on the Intelligence Index at 65.7 versus Astra’s 61.2, so this is not a clean sweep. It is also the first model OpenAI put at Critical cyber under the Preparedness Framework: ExploitBench 100% versus Sol’s 78.5%, two eval zero-days in the internal V8 port, and the Daybreak-gated offensive path is not the default product.

RSA-260 had sat unfactored since the challenge list was published, and today Eric Lu posted a 130-digit prime that divides it, with a cofactor that multiplies back to the published N, and no method. It is 260 decimal digits, 862 bits, the next challenge number after RSA-250 which fell in 2020.

Muse Spark 1.3 is $1.25 in and $4.25 out, the same Standard rate as 1.2, and the story is Meta's own claim of 20% fewer tool calls rather than another king on the Intelligence Index. It is live in Muse Code and on the Meta Model API, cache hits are $0.15. Meta also says about 25% fewer tokens versus 1.2, and that it asks before consequential actions. Artificial Analysis, separately, has the public xhigh at 61 and the partner-only max at 62, one under Opus 5.

Claude Fable 5.1 is not a leap from Opus 5 on the general index. Artificial Analysis has it at 66 at max effort, three points above Opus 5, and the knowledge-work evals sit inside the same noise. The leap is agentic scientific research. Terminal-Bench-Science 0.1 is the benchmark that asks a model to run real scientific workflows in a terminal, not to pass a quiz, and Fable 5.1 scores 52.6% on it against 29.0% for Opus 5 and 24.7% for Fable 5. Even with a standard error of about four points, that is almost double, which is the first time this generation actually moved on science agents rather than stacking another couple of points on the Intelligence Index.

OpenAI spent four years making models available through Cursor, then used a change-of-control clause the minute SpaceX bought them. They wrote a blog post that names Musk, cites X and SpaceXAI, and tells developers the editor they already live in is no longer trusted enough to get GPT. November 12 is the proposed shutoff. They called that the maximum notice the contract allows, which is a polite way of saying they used the smallest window they could still dress up as care. The line about caring deeply that models stay broadly available for developers is the joke. Broadly available, except in the coding tool a huge chunk of those developers actually open. Bring-your-own-key is the escape hatch they left so they can say they didn't cut anyone off, while they cut the native path that made Cursor useful in the first place. Anthropic is putting more Claude compute into the same editor. OpenAI is walking out of it and calling it safety.

The Information says Nvidia agreed to buy Hugging Face for $12.9B. Last year they wouldn't even take $500M at $7B valuation so Nvidia wouldn't own the table. Business Insider says the talks still have no signed paper and can still die. Neither company has said a word. The question isn't "did they buy the GitHub of AI". It's whether Hugging Face stays a public square once the chip company writes the check.

Claude Fable 5.1 is expected to release this week, with a high possibility of dropping today! > there is a chance it gets delayed to the start of the next week (per @synthwavedd) How do you think it will perform compared to the previous version? 🤔

Grok 4.6 xhigh scores lower than high on the Intelligence Index. Extra thinking bought more tokens, not a smarter model. Artificial Analysis: high is 61. xhigh is 60. Same $2 in / $6 out. On that eval, xhigh burned 75M output tokens. High burned 72M. High is the default. xhigh is the extra-effort knob. Turning it up made the model slower, wordier, and one point dumber. More effort is sometimes just bloat.

OpenAI didn't build a GPU. They built a chip that keeps KV cache local and refuses the latency tax. The underrated part: GPT helped design the thing. They used their own models to go from idea to tape-out in nine months. Codex + GPT-Astra programmed the chip so three extra open-weight families, not even in the original plan, hit high performance in two months. On selected GPT-OSS attention and MoE blocks, the AI-written kernels ran 1.5–1.8× faster than the human ones. That's not the full model. It is the loop: model designs chip, chip serves model, model writes faster kernels. Yesterday they put numbers on it. The models tested: GPT-OSS 120B, DeepSeek R1, Kimi K2.5 1T. 1.5–1.9× more work per watt than NVIDIA's GB200/GB300. 1.7–3.6× lower latency. Up to 4.1× on the interactive stuff agents actually need. DeepSeek goes 1.65s instead of 5.99s. Same request. Different silicon. openai.com/index/jalapeno-fi…

China's Zhipu AI just confirmed Ox Alpha is theirs. > new iteration of their GLM series > weights get released tonight. The model hit #1 on OpenRouter at zero cost and had more than double DeepSeek's usage. Biggest launch in that marketplace's history, still free for the rest of the week. It's a reasoning model for coding and agentic work. It takes text, image, and even video.

Grok couldn't code. That was the joke for a year. Fast on X. Fun at parties. Not the model you pointed at a repo. Then SpaceX bought Cursor. Not a chatbot. The IDE where the work already happens, plus the agent traces how engineers actually ship. Data you cannot scrape. Grok 4.5 jumped a lot because of that loop. 4.6 now ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index at $2 in, $6 out, and it lives in the editor they just bought. That's why this is one of the best buys of the cycle. Grok didn't get competitive again by posting more. It got competitive by owning the place code gets written. Name a better AI acquisition this year. I'll wait.

Why are we trying to slow down crucial progress for humanity? Altman and Amodei are the biggest pussies on this planet – at Bucharest, Romania

End of an era. Respect pentru cel mai mare fotbalist al tuturor timpurilor – at Bucharest, Romania

Estevao > Yamal

They can't get away with it forever

Lovitură liberă inexistentă la primul gol, ăla nu a fost penalty și acum cartonaș roșu nedat, rușine

Tocmai din cauza asta X e cea mai bună platformă de socializare. Nu ar trebui să existe cenzură și să stabilească un comitet ce e adevărat sau nu. E vina utilizatorului dacă crede orice știre falsă.

Regele