Anthropic released Fable 5.1 and Mythos 5.1 today, and the TLDR is: smarter, meaningfully cheaper to run, and, for the first time in this model line's short dramatic life, meaningfully less annoying about safety. The API id is claude-fable-5-1 and it's live on all platforms including AWS, Google Cloud and Azure.
Mythos 5.1 is the version with reduced safeguards for verified cybersecurity and life-sciences professionals, and Anthropic says that with today's safeguard improvements "we expect the difference between the models to be much smaller," since the old gap mostly measured tasks the imprecise classifiers were blocking.
The numbers
Terminal-Bench 4.0 agentic coding goes from 42.0 on Fable 5 to 55.8 on Fable 5.1, with Mythos 5.1 at 60.9. Agentic scientific research more than doubles, 24.7 to 52.6 on Terminal-Bench-Science. Humanity's Last Exam moves from 57.8 to 60.9 without tools, and CursorBench from 70.5 to 73.4.
Anthropic's flex example: Fable 5.1 "found the cause of a rare crash... that none of their engineers (or any other model) had been able to explain after several years."
Pricing stays $10 and $50 per million tokens, and the real cost news is caching: cache reads drop 75 percent to $0.25 per million, which Anthropic says cuts typical workloads by around 25 percent versus Fable 5 and up to around 45 percent on complex agentic ones.
The babysitter no more?
This is the part that everyone has been waiting for. After months of the classifiers flagging sunglasses questions and security homework, Anthropic is claiming concrete precision numbers: biology safeguards fire "85% less often for benign requests," and Claude Code users can expect "an average of around 60% fewer interventions per session" from the cyber safeguards. Fable 5.1 can now be used to discover software vulnerabilities, though not to develop exploits, which is roughly the defender-versus-attacker line I've argued the filters should draw.
What the announcement doesn't say
Nothing about subscription inclusion. No word on whether Fable 5.1 slots into the 50-percent-of-Max-limits arrangement Fable 5 ended up with after the July back-and-forth, no word on Pro, and no exact terms for Claude Code plans. Given that the last model's subscription saga ran for a month of extensions and reversals, that silence is the thing to watch.
Sources
Anthropic: Claude Fable 5.1 and Mythos 5.1 - The September 2026 release, the claude-fable-5-1 model id and all-platform availability, benchmark gains including Terminal-Bench 4.0 at 55.8 versus 42.0 and Terminal-Bench-Science at 52.6 versus 24.7, unchanged $10 and $50 pricing with cache reads cut 75 percent to $0.25 per million and typical workload costs down around 25 percent, the 85 percent reduction in benign biology flags and around 60 percent fewer cyber interventions per session in Claude Code, vulnerability discovery being allowed while exploit development is not, the continued redirection of dual-use tasks to Opus models, and the verification programs currently limited to US organizations.
