Anthropic introduced Fable 5.1 and Mythos 5.1 on September 1. It sounds like a small point release, but it’s the new flagship. And the two are technically identical: same model, just with different safeguards. Fable 5.1 is available to everyone, while Mythos 5.1 comes only through vetted access programs for cybersecurity and life sciences professionals, and for now only to US organizations.
Let’s start with price, because it hits most people right away. Fable 5.1 costs the same per token as Fable 5: ten dollars per million input, fifty per million output. But cache reads – the model re-reading context it has already processed – are now 75 percent cheaper, down to 25 cents per million tokens. For typical work that lands you at roughly 25 percent lower cost, and for heavily agentic tasks with lots of context, up to 45 percent. In other words, exactly the workloads that used to be expensive get noticeably cheaper.
On benchmarks the model pulls clearly ahead. On Terminal-Bench-Science, Fable 5.1 hits 52.6 percent, where Fable 5 sat at 24.7 and Opus 5 at 29. On Humanity’s Last Exam it’s 60.9 percent without tools, on CursorBench 73.4. Numbers are one thing. What I find more telling is the anecdote from the investment firm Millennium: an extremely rare crash, about one in a million runs, that nobody had been able to explain in four to five years, no human and no other model. Fable 5.1 disassembled a third-party vendor library, matched it against the core dump, and found the bug.
The part that surprised me most is the research. Mythos 5.1 designed protein binders whose binding affinity, on three targets, was ten times higher than the best entries in an open design competition. Its hit rate was almost 50 percent across twelve targets – today’s norm is 10 to 15. Fable 5.1 also trained a new elevation map for a third of the planet Venus from old Magellan radar data, which Anthropic is publishing under an open license. This isn’t a marketing demo; these are results you can read up on and reuse.
Safety moved as well. With Enterprise Frontier Safeguards, companies can keep their data on their own cloud, effectively zero data retention, rolling out this fall. The cyber safeguards now intervene around 60 percent less often, and Fable 5.1 is allowed to find software vulnerabilities for the first time, though not to build exploits for them. In biology, the filters fire 85 percent less often on benign questions.
My take: this is one of the rarer releases where more capability and lower cost arrive together instead of fighting each other. If you kept Opus 5 as your daily driver, you now have a very good reason to switch. And the research examples hint at where this is heading – a tool slowly turning into a colleague.
Sources: Anthropic: Introducing Claude Fable 5.1 and Mythos 5.1 · 9to5Mac · MacRumors