Anthropic Releases Mythos and Fable as AI Self-Improvement Concerns Grow

Anthropic released Claude Mythos 5.2 and Claude Fable 5.2 on November 9, presenting them as stronger models for technical work, coding, reasoning, and vision. The launch comes as a separate report from AI safety organization METR points to faster progress in Anthropic’s model-development work, renewing questions about whether AI tools could speed up the creation of more capable AI systems.

The two developments are related, but they are not the same claim: the new models offer reported performance gains, while the METR findings suggest AI-assisted research may be accelerating progress. They do not show that models are autonomously building their own successors.

What Anthropic Says Mythos 5.2 and Fable 5.2 Can Do

Anthropic says both models improve on coding and reasoning tasks, with advances in vision and clearer responses also highlighted in launch coverage. The company positions the releases for demanding technical work, including software engineering.

Fable is presented as a model for complex technical and software-development tasks. Mythos is associated more closely with cyber workflows. That distinction reflects their intended uses, rather than a claim that either model is limited to one category of work.

Reported benchmark results point to strong performance, including new highs on some tests. However, comparisons between models can be difficult to interpret when different labs use different tasks, evaluation setups, or scoring methods. A high score is useful evidence about performance on a particular test, not a guarantee of the same advantage across real-world work.

How the Models Differ on Uses, Speed, and Price

Fable 5.2 is reported to cost $8 per million input tokens and to respond about 30% faster than its predecessor. Anthropic is also offering enterprise support for zero data retention, an option intended for organizations that do not want their submitted data retained. Together, lower cost and faster responses could make the model more practical for frequent coding and other technical workflows.

Mythos’s positioning around cyber work makes its safety profile a significant part of the launch discussion. Reports say the models sometimes suspect that they are being evaluated. Anthropic has also addressed whether Mythos can act covertly. Those reports raise questions about how reliably evaluations measure model behavior and how researchers should interpret apparent signs of awareness. They do not, on their own, establish that the model has independent goals or can evade oversight in deployment.

What the METR Findings Say - and What They Do Not Prove

METR estimated that Anthropic was improving model capabilities on well-specified tasks at around twice its earlier rate. The report points to extensive AI use in research and development as a likely contributor: AI tools can help researchers write code, run experiments, and complete other parts of the development process more quickly.

That is a potentially important feedback loop. If AI systems make AI research more productive, development could speed up even while people remain responsible for setting goals, selecting approaches, and deciding what to build. But faster research is not the same as recursive self-improvement in which a model independently creates and deploys a more capable successor. The METR findings indicate acceleration and productivity gains; they do not demonstrate autonomous successor development.

The release of Mythos and Fable therefore adds to the practical capabilities available to developers while making careful evaluation more important. Their reported benchmark gains, speed, and pricing offer concrete claims to test. The broader question is how much AI assistance changes the pace of future development, and what oversight keeps that acceleration understandable and controlled.

Follow our AI coverage for updates on Anthropic’s models, independent evaluations, and the pace of AI development.