Anthropic calls the release its strongest system yet for coding and knowledge work. Beyond quality gains, it is trying to address three practical issues from the previous version: the cost of long sessions, excessive refusals on safe requests and business requirements around data handling.
One model, two boundaries
Fable 5.1 and Mythos 5.1 share the same technological foundation. The difference appears in the rules governing access to potentially dangerous capabilities.
Fable can look for vulnerabilities in source code, but it is not meant to help produce working exploits or carry out a complete intrusion. Mythos has looser restrictions for professional defensive work and biology research, but access is granted through dedicated vetting programs.
At launch, Mythos is available to a limited number of organizations in the United States. Anthropic says it will expand the program gradually, without setting out an international timetable.
✦ AILong agent tasks should get cheaper
Fable 5.1’s base price is unchanged: $10 per million input tokens and $50 per million output tokens. The lower cost comes from caching: rereading context that has already been processed now costs 75% less, or $0.25 per million tokens.
Anthropic estimates that ordinary workloads will cost roughly 25% less than Fable 5, while long agent workflows with a large amount of reused context could cost up to 45% less. Those are the developer’s estimates based on its own workload; actual savings depend on how a task is structured.
The change is especially relevant to Claude Code. An agent repeatedly rereads a project, its history and instructions. The larger the share of cached context, the more the new price matters.
Fewer false positives
Anthropic says its updated cybersecurity filters block 60% fewer safe requests. In biology tasks, the company says interventions in harmless scenarios have fallen even more.
That is an attempt to find a difficult balance. Weak protections expose dangerous capabilities; blunt protections turn an expensive professional model into a system that refuses to analyze ordinary code or a scientific question.
Independent coverage points to a trade-off. The system card says Mythos 5.1 is slightly more likely than Opus 5, in some evaluations, to accept a user’s unverified claim that they have authorization. That does not automatically make the model unsafe, but it explains the cautious rollout of its expanded capabilities.
Invisible text signals
Fable 5.1 adds statistical marking for generated text. It is not a visible label and contains no information about the user, but a dedicated interface can estimate how likely it is that Claude contributed to a piece of material.
Anthropic ties the feature to European regulatory requirements. Access to the detector is currently limited to regulators, researchers, news organizations and institutions that need it for legal compliance.
For an ordinary user, this means the text does not carry a visible watermark. The reliability of statistical detection remains an open question: it will need to be tested against edited, translated and mixed-source material.
What the tests show
Anthropic reports substantial gains in coding, scientific work and computer use. It also describes cases in which the model found a rare cause of failure or built a prototype during several days of autonomous work.
But most comparisons were run by the developer, and some test variants differ from earlier task sets. That means new percentages cannot always be compared directly with historical tables. Practical quality will be clearer after independent testing and broad use.
Bottom line
The central story in Fable 5.1 is not just higher scores. Anthropic has split one powerful model into a broadly available version and a professional version, trying to preserve usefulness without fully opening its riskiest functions. At the same time, cheaper caching makes long agent tasks more accessible. If the promised savings hold up, cost rather than a benchmark record may be the most visible change for developers.
