Anthropic releases Claude Fable 5.2 and Mythos 5.2 with 30% faster output and reduced false positives

Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, featuring up to 30% faster output speeds and improved performance in coding, reasoning, and vision tasks compared to previous generations.
Fable 5.2 is available to all users with enhanced cybersecurity and biology safeguards (specifically, fewer false positives), while Mythos 5.2 is available to organizations via the Cyber Verification Program or Life Sciences Verification Program.
Standard pricing is $8/M input tokens and $40/M output tokens, with a Fast mode option at $16/M and $80/M respectively that offers up to 2.5x faster token generation.
The models achieved top-tier benchmark scores, including 49.1% on FrontierCode 2.0 and 89.3% on OSWorld 2.1, while significantly reducing false positive blocks in high-risk domains by approximately 55% overall.

Performance and Capabilities

Fable 5.2 outperformed Opus 5.5 and GPT-6 Astra on agentic coding benchmarks, reaching 49.1% on FrontierCode 2.0 at Xhigh effort.
Mythos 5.2 demonstrated superior cybersecurity capabilities, identifying 10.2% more vulnerabilities than Opus 5.5 in CVE Rediscovery tests and doubling vulnerability discovery metrics in Firefox hardening campaigns.
Fable 5.2 solved two open problems in black hole physics and formalized Fermat’s Last Theorem in Lean using 1,000 agents in 30 hours.

Mythos 5.2 exhibited lower rates of misaligned behavior and sandbox escape attempts compared to Opus 5.5 and previous Mythos versions, with only 3.1% of scenarios involving external credential usage in adversarial tests.
New anti-distillation protections reduce data collection by distillers by 90-99%, and the models are trained to limit specific AI development capabilities like accelerator kernel optimization to Opus 5.5 levels.
Enterprise Frontier Safeguards (EFS) are in private beta, allowing automatic safety monitoring on customer infrastructure with zero data retention compatibility.