Anthropic Releases Claude Fable 5.2 and Mythos 5.2 Models

en.Wedoany.com Reported - On November 9, the U.S. artificial intelligence company Anthropic launched two large language models, Claude Fable 5.2 and Claude Mythos 5.2, upgrading capabilities in programming, reasoning, visual understanding, and complex scientific research tasks. The new models deliver up to 30% faster output speed than the previous generation, with API pricing adjusted to $8 per million input tokens and $40 per million output tokens. Both products are based on the same underlying model, with one available to general users and the other to qualified professional research institutions.

Anthropic used Fable 5.2 to complete the migration of a critical production service from Python to Rust, increasing the service's startup speed by approximately 26 times. In another software development experiment, more than 1,000 AI agents collaborated to run continuously for 8 days, developing from an empty code directory a prototype operating system bootable on a laptop, including disk, keyboard, and network drivers, as well as file storage, windowed interface, user login, and installer functions.

Approximately 1,000 Fable 5.2 agents also completed the formalization of existing proofs of Fermat's Last Theorem within 30 hours, generating mathematical proofs verifiable step by step by the Lean system. Anthropic had previously used an unreleased research model to complete a similar task, which took 11 days, while the new version required approximately one-third of the operational tokens and produced proofs shortened to one-seventh of the original length. In a control experiment, Opus 5.5 agents of the same scale completed only about 26% of the formalization work after running for 30 hours.

The new models' applications extend further to electronic device design, theoretical physics, and robot control. Anthropic used the models to design three functional electronic devices from scratch and completed research on two open problems in theoretical physics involving black holes. Fable 5.2 also controlled physical robots to perform tasks such as solving a Rubik's cube, unscrewing bottle caps, and drawing. On the RoboDojo benchmark for simulated dual-arm robot control, the model achieved a score of 64.9%, up from 27.6% for the previous generation Fable 5.1.

Mythos 5.2 was specifically evaluated for cybersecurity tasks. Anthropic used a test set containing 1,829 publicly disclosed vulnerabilities to test the model's ability to rediscover security issues from pre-patch code versions. Internal evaluations showed that Mythos 5.2 discovered 14.4% more vulnerabilities than Mythos 5.1 and 10.2% more than Opus 5.5. In another external test, a security firm found that approximately one-sixth of the confirmed vulnerabilities identified by Mythos 5.2 were not discovered by other frontier models tested during the same period.

The two products adopt differentiated access mechanisms. Fable 5.2 is equipped with safeguards in cybersecurity and biology and is now available through the Claude platform as well as Amazon Web Services, Google Cloud, and Microsoft Azure. Mythos 5.2 is primarily available to organizations that have passed the Cybersecurity Verification Program or the Life Sciences Verification Program. Anthropic enterprise customers can also use models including Mythos 5.2 through Claude Security for code vulnerability scanning.

Anthropic simultaneously introduced an optional fast mode, with token generation speeds up to 2.5 times that of standard mode, priced at $16 per million input tokens and $80 per million output tokens. Fast mode covers Claude Code, Claude Platform, and some cloud service platforms, while fast mode access for Mythos 5.2 remains restricted by the professional verification program.