Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2，强化科研与编码能力

Anthropic PBC today pulled back the curtains on Claude Fable 5.2 and Claude Mythos 5.2, its two most capable large language models to date.

The LLMs are identical as far as their core capabilities are concerned. The difference is that Fable 5.2 blocks certain biology and cybersecurity use cases, while Mythos 5.2 includes more relaxed guardrails. The latter model is only available to trusted researchers through a pair of limited-access programs.

Anthropic tested Fable 5.2 against its predecessor across 14 artificial intelligence benchmarks. The model beat Fable 5.1’s score by more than 10% across most of the tests. It’s particularly adept at scientific research: Fable 5.2 scored 74.6% on Terminal-Bench-Science 0.1, a 22% improvement over Fable 5.1.

Earlier this year, Anthropic used an unreleased LLM to generate a computer-verifiable version of a famous mathematical proof. The project took 11 days. According to Anthropic, Fable 5.2 reduced the workflow to 30 hours by spinning up 1,000 agents to parallelize calculations. In a separate project, it answered two open questions about a theoretical physics concept related to black holes.

The company also tested Mythos 5.2, the invite-only version of Fable 5.2. It asked the model to study the protective shells of harmless viruses called AAVs. AAV shells, or capsids as they’re known in scientific parlance, are used by some therapies to deliver drugs.

Mythos 5.2 accurately modeled the structure of 28 out of the 34 AAV capsids it studied. The result is notable because capsids comprise dozens of different proteins. Other cutting-edge models that are optimized for biology research, notably Google DeepMind’s AlphaFold2, can only model the structure of individual proteins. The creators of the latter model won one half of the 2024 Nobel Prize in chemistry.

Fable 5.2 includes guardrails that block most research biology questions and many cybersecurity-related prompts. If users enter a request that falls into either of those categories, the model sends it to Opus 4.8 or Opus 5. Anthropic says Fable 5.2 is 55% less likely to erroneously block safe prompts than Fable 5.1.

The model outperforms its predecessor by more than 13% across two coding benchmarks called FrontierCode 2.0 FrontierSWE 2.0. According to Anthropic, one of its engineers put the model’s programming capabilities to the test by asking it to build an operating system. Fable 5.2 spun up up 1,000 agents that completed the task in about eight days. The operating system they created can run on a laptop, offers network access and features a Claude sidebar.

Anthropic sees customers using Fable 5.2 and Mythos 5.2 to speed up highly complicated, open-ended technical tasks. The company says more day-to-day workflows can be automated more efficiently using Opus 5.5, an LLM it introduced in July. Users can also combine the models by having Opus 5.5 ask Fable 5.2 or Mythos 5.2 for advise when it encounters a bottleneck.

Fable 5.2 and Mythos 5.2 are priced at $8 per million input tokens and $40 per million output tokens. That makes them 20% more affordable than Anthropic’s previous-generation models. Both LLMs offer a high-performance mode that generates prompt responses 2.5 times faster than the standard tier for double the price.

Image: Anthropic