Anthropic launches Fable 5.2 and Mythos 5.2 with science, coding upgrades

Anthropic unveiled Claude Fable 5.2 and Claude Mythos 5.2, its most advanced language models to date. The models are identical in core capabilities, with the exception that Fable 5.2 restricts certain biology and cybersecurity use cases, whereas Mythos 5.2 has looser guardrails. Only Mythos 5.2 is accessible to select researchers through limited programs.

Compared to its predecessor, Fable 5.2 improved by over 10% in 14 AI benchmarks, excelling particularly in scientific research. For instance, it scored 74.6% on Terminal-Bench-Science 0.1, a 22% improvement over the previous version. Anthropic utilized an unreleased LLM to generate a proof of a famous mathematical theorem, cutting the workflow from 11 days to 30 hours by deploying 1,000 agents to parallelize calculations.

Mythos 5.2 was also tested, which accurately modeled the structure of 28 out of 34 AAV capsids, a feat that other biology-focused models struggle with. The model's guardrails block most biology and cybersecurity-related prompts, making it 55% less likely to mistakenly block safe queries. Fable 5.2 outperformed its predecessor by 13% in two coding benchmarks.

One engineer tested the model's programming skills by asking it to create an operating system, which was completed by 1,000 agents in about eight days. The operating system can run on a laptop, offers network access, and features a Claude sidebar. Anthropic envisions customers using these models to expedite complex, open-ended technical tasks, while Opus 5.5, introduced earlier this year, can automate more routine workflows.

Both models are priced at $8 per million input tokens and $40 per million output tokens, making them 20% cheaper than Anthropic's previous-gen models. They also offer a high-performance mode that generates responses 2.5 times faster than the standard tier at double the price.