Anthropic is developing a new AI model with sharply improved performance compared with its existing models and has begun testing it with some customers, Fortune reported on March 26.

The development became known after Anthropic mistakenly exposed an unpublished document in a public data repository. Fortune said it first confirmed the document and reported the details.

According to the unpublished document, an Anthropic spokesperson said the new model has achieved a “step change” in AI performance rather than an incremental improvement and shows the best performance among models developed so far.

Anthropic is currently running trial operations with a small number of early-access customers.

After being notified by Fortune of the data leak, Anthropic blocked external searching and document viewing of the data repository.

The leaked document listed the new model name as “Claude Mythos.” The same document also mentioned a new model tier called “Capybara.” It said Capybara is a larger, more intelligent tier above the current top tier, Opus, and is more expensive. Mythos and Capybara appear to refer to the same base model.

Anthropic currently releases models in three sizes: Opus, Sonnet and Haiku. Capybara is a new tier added above them.

According to the document, the new model scored far higher than the existing top model, Claude Opus 4.6, across multiple evaluations including software coding, academic reasoning and cybersecurity. Responding to media inquiries about the new model, Anthropic acknowledged it is developing a general-purpose model that has made meaningful progress in reasoning, coding and cybersecurity.