Anthropic launches Claude Fable 5.2 and Mythos 5.2 with 30 percent faster output and lower pricing

Anthropicrecommends starting with Claude Opus 5.5 for coding and general knowledge tasks before moving to Fable or Mythos for the most challenging open-ended technical work. Fable 5.2 can also serve as an advisor to smaller models like Haiku, Sonnet, or Opus when they encounter difficult problems, or as an orchestrator that distributes simpler tasks to smaller models.

Both models cost 8 dollars per 1 million input tokens and 40 dollars per 1 million output tokens. A fast mode that generates tokens up to 2.5 times faster costs 16 dollars and 80 dollars respectively. On the FrontierCode 2.0 benchmark at the highest difficulty level, Fable 5.2 scored 49.1 percent, exceeding Opus 5.5 at 45.5 percent and GPT-6 Astra at 42.8 percent, while using roughly half the cost per task. The model also achieved 69.9 percent on FrontierSWE v2, 89.3 percent on OSWorld 2.1, and 74.2 percent on Humanity's Last Exam with tools.

Fable 5.2 is now available on all platforms including Amazon Web Services, Google Cloud, and Microsoft Azure, accessible through the Claude Platform using the model identifier claude-fable-5-2. During early testing, an agent team running Fable 5.2 migrated a critical Anthropic production service from Python to Rust, achieving a startup speedup of approximately 26 times. In another experiment, the model coordinated over 1,000 subagents for 8 days to build a prototype operating system from an empty folder, resulting in a bootable system with drivers for disk, keyboard, network, file system, windows, and login page.

Mathematical agents using Fable 5.2 formalized the proof of Fermat's Last Theorem in the Lean language within 30 hours, compared to 11 days in Anthropic's first attempt. Fable 5.2 also solved 2 open problems in theoretical physics related to black holes, designed 3 functional electronic devices with Diode, and scored 64.9 percent on the RoboDojo Sim robot control benchmark, more than double Fable 5.1's 27.6 percent.

In cybersecurity testing, Claude Mythos 5.2 discovered 10.2 percent more vulnerabilities than Claude Opus 5.5 and 14.4 percent more than Mythos 5.1 in Anthropic's internal CVE Rediscovery test measuring 1,829 published vulnerabilities. Mozilla found that Mythos 5.2 discovered roughly twice as many Firefox vulnerabilities as Mythos 5.1. Claude Enterprise customers can access Mythos 5.2 through Claude Security even without Cyber Verification Program membership, which scans codebases for vulnerabilities. Fable 5.2 blocks high-risk cybersecurity work such as penetration testing and exploit creation, as well as most biology research, routing these requests to Opus 4.8 for cybersecurity and Opus 5 for biology instead. Anthropic reduced false positives in its safety classifier by approximately 55 percent compared to Fable 5.1.

For safety and alignment, Anthropic assessed that the risk of unsafe model behavior in critical situations remains low. In adversarial sandbox escape testing across over 1,800 scenarios, Mythos 5.2 used credentials or files outside the sandbox in 3.1 percent of cases, comparable to Opus 5.5 at 3.3 percent and significantly better than Mythos 5.1 at 21 percent. Fable 5.2 includes distillation protection that reduces information malicious actors can extract by 90 to 99 percent. The preserved thinking feature in the application programming interface now extends to all accounts. Organizations requiring zero data retention can use Enterprise Frontier Safeguards, currently in private beta, which runs automated safety checks on prompts and responses stored on the customer's own infrastructure. Other customers have data retained for 30 days for safety verification.