Anthropic PBC today introduced Claude Mythos 5 and Claude Fable 5, two large language models that it says outperform the competition across a wide range of benchmarks.

The LLMs are derived from the Claude Mythos Preview algorithm that the company debuted in April. The model made headlines for its ability to find highly complicated cybersecurity vulnerabilities. According to Anthropic, Mythos 5 and Fable 5 both surpass its performance.

Fable 5 is broadly available to the public. It blocks prompts that relate to high-risk use cases such as cybersecurity vulnerability discovery. Mythos 5, by contrast, includes more relaxed guardrails and will be available only to a limited number of organizations. Anthropic PBC will work with the U.S. government to manage access to the model. 

The company says that Mythos 5 is the first model in the Claude series with the ability to “consistently produce novel, compelling scientific hypotheses.” The company asked the LLM to propose explanations for several molecular biology phenomena that scientists don’t yet fully understand. According to the company, several of Mythos 5’s suggestions were so promising that its researchers decided to launch a series of experiments aimed at verifying them.

One of the model’s hypotheses has already been corroborated in a lab. According to Anthropic, the model discovered new information about one of the proteins that make up the E. coli bacterium.

Another internal test saw Mythos 5 discover 14 protein targets, biological building blocks that can potentially be used to make medicines. Anthropic claims that nine of the targets proved to be “strong candidates for drug design.” Furthermore, Mythos 5 speeded up some of the tasks involved in the protein discovery process by a factor of 10.

Fable 5, Anthropic’s other new LLM , blocks requests that relate to cybersecurity, biology and chemistry. The built-in guardrails reroute such prompts to Claude Opus 4.8, a less advanced model that Anthropic launched in May.

Both Fable 5 and Mythos 5 have set a record on the SWE-Bench Pro programming benchmark with a 80.3% score. Stripe Inc., an early adopter of the former model, used it to modernize an internal software repository with 50 million lines of code. Anthropic says that Fable 5 enabled the company to complete the task in a day instead of the two months it would have otherwise required.

Fable 5 and Mythos 5 also provide significant performance gains across several nontechnical use cases. They scored 7.3% higher than Opus 4.8 on the GDP.pdf benchmark, which comprises document review tasks. Additionally, the models set a record on a second benchmark that measures LLMs’ ability to automate legal tasks.

Anthropic’s new models are priced at $10 per million input tokens and $50 per million output tokens. That’s less than half what the company charged for Mythos preview. Furthermore, Mythos 5 and Fable 5 use fewer prompts to complete tasks.

Image: Anthropic