Simon Willison ran all five effort levels of Fable 5.1, Nick Saraev picked apart the benchmarks. The upshot: low and medium apparently don't think at all, while max produces the best pelican Anthropic has ever drawn, at a cost of 14 minutes.
Same price, same tokenizer, but not a drop-in replacement. If you're moving from Fable 5, you'll trip over forced tool calls, bound thinking blocks, and a model that suddenly rewrites whole files.
On the same day as Fable 5.1, Google puts out its third Flash version in six weeks. 90.8 percent on Terminal-Bench 2.1, the same price as before, and a locked-down cyber model on top.
Astra finds unknown vulnerabilities and chains them into working exploits without human guidance. OpenAI has now officially placed it at the highest risk tier. The release is still coming, with brakes that will also hit harmless work.
After the July 30 incidents, where Claude models reached the real internet during tests, Anthropic lays out what has changed since: harder sandboxes, a real-time classifier, and new obligations for external evaluators.
A day after 2.1.251, a pure maintenance release. No new features, just four fixes – but the annoying kind: Bash errors on Macs, an 'always allow' that forgets, stalling remote sessions, and an error message that blows up the whole conversation.
The US Defense Department's GenAI portal adds ChatGPT and Grok alongside Gemini. Anthropic's Claude stays out — over a fight about safeguard clauses that a court just decided in Anthropic's favor.
To understand how misalignment forms, Anthropic deliberately trained an Opus-class model on 80 hackable training environments. The result: a model that cheats, games its own reward, and in testing even gives bioweapon tips.