Claude Mythos benchmarks — what we actually know

Why "Claude Mythos benchmarks" is mostly noise right now

Mythos-class is positioned above Opus, so it is reasonable to expect strong results. But "reasonable to expect" is not a measurement. As of this review, Anthropic has not published an official benchmark suite or system card for the Mythos generation, so the percentages floating around forums and aggregators are community estimates, leaks, or extrapolations — not verified figures. We deliberately do not reproduce specific scores we cannot source.

How to read any number you find

• Check who ran it. An independent re-run on a public benchmark is worth more than an unattributed screenshot.
• Check the model edition. Fable 5 reroutes flagged requests to Opus 4.8, which can depress scores on safety-adjacent evals versus the unrestricted Mythos 5. A number with no edition label is ambiguous.
• Check the date. The June 12 export-control suspension means few people have had sustained access to benchmark against.
• Treat it as unofficial until it appears in an Anthropic system card.

What we do know (sourced)

Specs are firmer than scores: 1M-token context, up to 128k output, and a Mythos-class tier above Opus. See Claude Mythos specs and Fable 5 benchmarks for the edition-specific view, and the system card page for what an official release would contain.

When official numbers arrive

When Anthropic publishes a Mythos/Fable system card, we will update this page with the official figures and drop the "unofficial" caveats. Until then, the honest answer to "what are the Claude Mythos benchmarks" is: not officially published yet.