Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Sept. 1, 2026. Anthropic says the two are the same model with different levels of safeguards: Fable 5.1 is generally available, while Mythos 5.1 is limited to trusted access programs for cybersecurity and life sciences work. That split lets Anthropic sell one model widely while keeping its least restricted form for vetted groups.
Fable 5.1 keeps its predecessor's prices: $10 for every million chunks of text it reads and $50 for every million it writes, where a chunk is often part of a word. Reusing text the model has already been sent, known as a cache read, now costs $0.25 per million, which Anthropic describes as a 75% cut from Fable 5's $1.00. Bulk jobs sent through Anthropic's batch service stay half price. The model can take in up to 1 million chunks of text at once, its knowledge runs to June 2026, and developers call it claude-fable-5-1.
In Anthropic's own testing, Fable 5.1 scored 52.6% on a test of scientific research tasks that an AI carries out on its own. In the testing company Vals.ai's public results for the same test, it scored 34.29%, about half the 65.71% of OpenAI's GPT-6 Astra. Anthropic notes that the safeguards on its released model can give some tasks a score of zero, and different setups, numbers of tries and effort settings can move scores on the same test.
The testing firm Artificial Analysis gave Fable 5.1, set to maximum effort, a score of 66 on its Intelligence Index, a combined score built from many tests, in a Sept. 1 review. As of Sept. 14, 2026, its live page showed 53 after a revision of the index, a change in the yardstick rather than in the model. The Sept. 1 review also found that running the index cost about 20% more with Fable 5.1 than with Fable 5, because the new model produced about 1.7 times as much output, despite the cheaper cache reads.
Anthropic also reports science results. It says Mythos 5.1 designed protein binders, proteins made to latch onto a chosen target, that outside organizations tested, with a hit rate of about 50% across 12 targets, and that on three targets from an Adaptyv Bio design competition its designs bound about 10 times more strongly than the best earlier entries. It says Fable 5.1 built a more detailed height map of Venus from Magellan radar data, released under a Creative Commons license, and that Mythos 5.1 sped up seven open-source biology models by up to 2.5 times, with the faster code to be released openly. No outside lab has yet published its own account of how these results were produced.
Anthropic's system card, the company's published overview of how a model was built and tested, says Fable 5.1 and Mythos 5.1 are the same trained model. Fable keeps blocks on biology and cyber work that could serve harm as well as good, while Mythos relaxes them for vetted programs. The card rates the model at CB-1 for chemical and biological risk, short of CB-2, keeps Fable on the same biology safeguards as Fable 5, and raises Anthropic's estimate of the risk of catastrophic harm from the model acting against what its makers intend, from very low to low, after recent cyber-evaluation incidents.
Fable 5.1 is allowed to find vulnerabilities, weaknesses in software that an attacker could use, in source code at every access level, while requests to build attacks and several other cyber tasks that can serve both defense and harm are still redirected. Anthropic also pitches Enterprise Frontier Safeguards, which give business customers control of monitoring data, and offers eligible customers zero data retention until those safeguards roll out.