Anthropic released its latest high-end model, Claude Fable 5.2, on Monday, and simultaneously launched Claude Mythos 5.2, which is limited to professional institutions and focuses on large-scale program development, scientific research, and long-term autonomous execution tasks. Fable 5.2 scored 65 points in the latest Intelligence Index of the third-party evaluation agency Artificial Analysis, ranking first in the comprehensive rankings. Anthropic also lowered the API prices of the two models, charging $8 and $40 per million input and output tokens respectively, which is 20% cheaper than the previous generation.
One of the most watched results published this time was the use of about a thousand Fable 5.2 AI agents to complete the formal proof of Fermat’s Last Theorem in 30 hours. Fermat's last theorem has long been proven by mathematicians, and formalization involves converting existing mathematical derivation into something that computers can check step by step. Anthropic used an internal research model to complete the same work in September this year, which took 11 days and required specially designed verification tools and researcher guidance. This time it was handled by a large number of agents, so the level of manual intervention was greatly reduced, and the proofs produced were much simpler than the previous time.
In terms of software engineering, Anthropic said that Fable 5.2 has helped internal teams handle large projects that were originally too costly and difficult to promote. One of the tests migrated important formal environment services from Python to Rust, and the completed system startup speed increased by about 26 times. Another experiment allowed the model to coordinate more than 1,000 subagents, and it took 8 days to create an operating system prototype that could be launched on a laptop from a blank folder, including file storage, keyboard, network, window interface, and login functions. Anthropic also uses Fable 5.2 for hardware and theoretical physics research. The team cooperated with Diode, a circuit board design company, and let the model be responsible for selecting electronic components, designing circuit boards, planning casings, and writing software. Finally, three operational devices were produced: a voice assistant controller, a USB conferencing device, and a handheld electronic pet. The manufacturing and assembly of physical circuit boards are still participated by partner manufacturers and are not completely completed by AI independently.
In terms of scientific research, Anthropic claims that Fable 5.2 helps solve two unsolved theoretical problems related to black holes and quantum gravity. One involves the formation of black holes under specific quantum conditions, while the other proposes more precise limits on a long-standing method of analyzing quantum gravity theory. Relevant results have been compiled into research preprints, but they still need to undergo subsequent academic testing.
Model reviews also show improvements in the new version. According to data released by Anthropic, Fable 5.2 achieved 49.1% in the FrontierCode 2.0 agent program development test, which is higher than Claude Opus 5.5's 45.5%, OpenAI GPT-6 Astra's 42.8%, and Google Gemini 4 Carbon's 41.1%. In the FrontierSWE v2 software engineering evaluation, Fable 5.2 also led with 69.9%. However, Google Gemini 4 Carbon achieved 53.9% in AutomationBench, which is still higher than Fable 5.2's 51.4%. The performance of each model is not consistent on different tasks. The third-party Artificial Analysis gave Fable 5.2 a highest overall score of 65 points, surpassing the previously leading Gemini 4 Carbon. However, high inference settings usually require more tokens to be consumed, and the actual cost of use cannot be judged solely based on the ranking scores. Anthropic also recommends that most general program development and knowledge work can still use Opus 5.5 first, and then consider switching to Fable when encountering particularly complex tasks that require long-term planning.
In terms of price, both Fable 5.2 and Mythos 5.2 have dropped from US$10 per million inputs and US$50 per million outputs to US$8 and US$40 respectively. Anthropic claims that the new version's text generation speed is increased by up to 30%, and it also provides an optional fast mode, which can be increased by about 2.5 times at the fastest, but the input and output rates have also increased to US$16 and US$80 respectively. Both models support 1 million token contexts, with a maximum single output of 128,000 tokens. The deadline for training knowledge is June 2026.
Fable uses the same underlying model as Mythos, with the main difference being restrictions on high-risk uses. Fable 5.2 will restrict requirements for exploit development, some penetration testing, and sensitive biological research, and is suitable for general professional work; Mythos 5.2 will open up more information security and life science research capabilities, but it must pass Anthropic's identity and purpose verification. The latter has begun to be available to qualified institutions for application, and is not a version that general subscribers can directly choose. Anthropic said Mythos 5.2 identified more known security issues than its predecessor in internal vulnerability rediscovery testing. Mozilla also participated in testing the new version, and the Firefox engineering team reported that the new model has significantly improved vulnerability research compared to Mythos 5.1. General enterprise customers can use Claude Security to use model capabilities including Mythos 5.2 to perform code security scanning.
Claude Fable 5.2 has been launched on November 9, supporting qualified Claude payment plans, Claude Code and Claude API. Developers can use the model identification code "claude-fable-5-2" to connect. Amazon Web Services, Google Cloud and Microsoft Azure are also included in the provisioning platform. Anthropic also provides limited-time additional usage credits for some paying subscribers. The actual available amount varies depending on the subscription level and account qualifications.
