The Capybara in the Room

Updated 20 September 2026: this article now distinguishes the early Mythos and Capybara naming discussion from Anthropic’s subsequent public releases. A codename is not evidence of a separate shipping product, a release date, or a capability ranking.

From Codenames to Public Evidence

The useful engineering question is what a model can do under stated conditions, and whether a particular team can use it. Rumors about names, leaked repositories, investor reactions or launch odds cannot answer that question.

Anthropic’s 7 April 2026 Mythos Preview research announcement described a restricted release focused on cybersecurity work. The company reported results from its own evaluations and explained why it was limiting access. Those are attributed vendor findings, not an independent demonstration that the model is best at every task.

A security benchmark needs its task definition, tooling, attempt budget and scoring method beside the headline result. Percentages from different benchmark variants or different scaffolds are not automatically comparable. A successful vulnerability demonstration also does not establish reliability across an entire software estate.

Access Is Part of the Product

Anthropic’s June announcement of Fable 5 and Mythos 5 documents a later stage of the product line. Its current Mythos page should be checked for present availability. The original naming discussion should not be read as current purchasing guidance.

For an engineering team, availability includes the account and deployment route, tool access, operational limits and the terms that apply to that route. A research announcement alone does not establish a public API price or a general release.

What to Evaluate

Use a representative task set with explicit acceptance criteria. Record the model identifier, date, tools, context, time and cost, then review failures as closely as successes. For security work, operate within an authorized test scope and verify findings through the project’s established reporting process.

The argument for paying attention to a new capability tier is practical: stronger models may change which engineering tasks are economical to delegate. That possibility deserves evaluation. It does not require unsupported claims about stock-market motives, legal confidence, or a release that changes everything.