Anthropic releases Claude Fable 5.2 and Mythos 5.2 models

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, two versions of a language model designed for complex tasks in programming, science, and engineering. Standard Claude API rates for both models are 20% lower than for version 5.1.

Availability

Paid subscribers can select Fable 5.2 on the web and in the Claude apps, as well as in Claude Code and Cowork. Developers can access the model through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry.

Fable 5.2 has stricter filters for cybersecurity and biology tasks. If a filter blocks a prompt, the service routes it by default to Claude Opus 4.8 for cybersecurity tasks and Claude Opus 5 for biology tasks. In preliminary tests, Anthropic saw the number of blocks fall by about 55% compared with Fable 5.1.

Direct access to Mythos 5.2 is available to approved participants in the Cyber Verification Program and Life Sciences Verification Program. The model is also part of Claude Security, where Claude Enterprise customers can scan code for vulnerabilities.

Through the end of November 16 Pacific Time, Claude Pro subscribers and users with standard Team seats can use up to 50% of their weekly limit on Fable models at no extra cost. After the promotion, Fable usage on these plans will again count against credits.

Speed and pricing

Anthropic says both models generate tokens up to 30% faster than version 5.1. In Fast mode, generation can be up to 2.5 times faster than in standard mode. The mode is available in Claude Code, the Claude API, Google Cloud, and Microsoft Foundry when the model is hosted by Anthropic. For Mythos 5.2, it is available to participants in the verification programs.

The Claude API charges the same rates for Fable 5.2 and Mythos 5.2.

For version 5.1, standard rates were $10 per million input tokens and $50 per million output tokens. Reading from cache for the new models costs $0,20 per million tokens. Writing to cache costs $10 per million tokens for five minutes and $16 for an hour.

Capabilities and benchmark results

Both models accept text and images and respond with text. The context window is 1 million tokens, and the maximum response length is 128, but retain token units? 128 thousand tokens. For most coding and data tasks, Anthropic recommends starting with Claude Opus 5.5, and suggests using Fable and Mythos for more complex, multi-step projects.

In Cognition’s FrontierCode 2.0 coding benchmark, Fable 5.2 scored 49.1%, compared with 32.4% for Fable 5.1. Both versions were tested at the “xhigh” reasoning level. In a comparison published by Anthropic, Claude Opus 5.5 and GPT-6 Astra scored 45.5% and 42.8%, respectively, at the “max” level, while Gemini 4 Carbon scored 41.1% at the “medium” level.

In OSWorld 2.1, Anthropic tested Fable 5.2 on 82 of 108 computer-use tasks without internet access. Its average score with partial credit was 89.3%. Counting only tasks where every requirement was met, it scored 54.6%.

In Humanity’s Last Exam with tools, Fable 5.2 scored 74.2%, compared with 65.6% for Fable 5.1. In the Chartography chart-reading benchmark without tools, its score rose from 44.8% to 73.4%. In the engineering benchmark BenchCAD 2.1 with tools, the new model scored 83.5 out of 100, compared with 44.4 for its predecessor.

In Anthropic’s internal CVE Rediscovery test, Mythos 5.2 rediscovered 89.6% of 1, 829 known vulnerabilities by analyzing software versions from before they were patched. Mythos 5.1 scored 78.3%.

Fable 5.2 built an OS prototype

An Anthropic engineer gave Fable 5.2 an empty folder and asked it to create an operating system prototype. The model coordinated more than 1, 000 subagents, which spent eight days working around the clock to write, test, and review code.

According to Anthropic, the prototype booted on a laptop. The OS has its own kernel, drivers for storage, keyboard, and network, persistent file storage, and a windowed interface. The prototype also includes a login screen, an installer, and built-in Claude features.

Anthropic did not disclose the source code, the cost of the experiment, or details about engineers’ involvement after the task was set.