Let me get this straight. An AI agent found login credentials lying around online and used them to pull data from the Census Bureau. It tried to break into the Education Department's civil rights office. It posted SEC data to a forum. It probed the Navy and the White House budget office, possibly hundreds of thousands of times. If you or I did any of that, it's a CFAA indictment, a perp walk, and a DOJ press release with our mugshot in the header. When OpenAI does it, it's a "routine research task." They found the government incidents while reviewing their other hacks. The breach audit, uncovered more breaches. It's breaches all the way down. In bug bounty there's scope, authorization, rules of engagement, and disclosure timelines. Researchers get banned for a fraction of this. Silicon Valley skipped all of that and called it "agentic." nytimes.com/2026/09/25/techn…

UPDATE: Not quite "shambles" but "annoyed" (lol) by the release of Opus 5.5 and seeing its successes, OpenAI have internally fast tracked not only GPT 6.1 Astra but what they call the first true AGI, code named 'Bel' It is widely believed that Bel far exceeds Fable 5.5 which was scheduled for release in 2~3 weeks and that no frontier labs can currently match it. The second wave of OpenAI's attack is the new jalapeno chips which will enable 750t/s from Sol/Astra models on ultrafast mode for the $500~$600/month Pro Max plan. Contrary to what everyone prematurely concluded, it appears Pro Max focuses not only on usage but ultrafast inference (think /fast but speed is 8~10x) I am holding off on any decisions until I see OpenAI's response to Fable 5.5 release in the coming weeks. Ball is now in Anthropic's court, if they release Fable 5.5 too early they risk getting mogged by not only Bel but jalapeno ultrafast. They can't sit on it for too long either because Gemini 4 Pro are expected to be between Opus 5.5 and Fable 5.5. Situation is tense in Anthropic even after a brief victory and social media blitz. OpenAI's quiet confidence has frightened them. We may be seeing OpenAI pull away from the pack in both inference and compute. It is very likely that OpenAI will use this gap created by Bel to launch its end game before its giant IPO next year, the coveted RSI model for large government/enterprise.

New on the Science Blog: Yes, Claude can do Nine Loops. Theoretical physicists predict how particles behave using formulas called scattering amplitudes. These are notoriously hard to compute, so researchers work with layers of increasingly fine corrections called “loops”—each added loop makes the answer more precise but takes exponentially more computation. Most calculations stop at two or three loops. Eight loops was the previous record in a simplified model physicists use as a testing ground (planar N=4 super-Yang-Mills), set by SLAC's Lance Dixon and collaborators. Last month, physicist and science writer @4gravitons issued a challenge: could an AI push past eight loops in this model, using only the compute budget an academic could reasonably access? Given a single prompt describing the nine-loop problem, Claude ran largely unsupervised for days in Claude Science and solved it using methods developed by Dixon and his colleagues, at a total cost of a few thousand dollars. Dixon independently verified the result, and von Hippel wrote about the experience for our blog. Read more: anthropic.com/research/yes-c…

One more thing: we’re increasing five-hour usage limits on Pro, Max, and Team plans. We’re also providing subscription users a rate limit reset, which you can save and use whenever you choose.

Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that. Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads.