anthropic/ large-language-models · ai-safety

Anthropic releases Claude Fable 5 with top benchmark scores

On June 9, 2026 Anthropic launched Claude Fable 5, a public Mythos model that tops MMLU, HELM and other benchmarks while being billed as safe for general use.

Anthropic releases Claude Fable 5 with top benchmark scores
  • Anthropic made its Mythos model publicly available as Claude Fable 5.

Claude Fable 5 arrived on June 9, 2026 as the first public release of the Mythos family. Anthropic says the model is “safe for general use” after internal evaluations that flag reduced toxic output and fewer hallucinations. In testing it hit the highest scores on the MMLU, HELM and BIG‑bench suites, surpassing the previous Claude 3.5 release by a few points on each metric. The company also posted a 2 × speed improvement on multi‑core hardware, citing a 1.8‑second latency for a 2 k token prompt.

The claim matters because most new frontier models stay behind paywalls, limiting independent verification. By opening Fable 5, Anthropic gives researchers a baseline for comparing safety‑focused LLMs. Still, “safe for general use” is a marketing label that hinges on proprietary tests, not an independent audit.

The rollout puts Anthropic back in the public‑model race, but the safety badge remains a moving target.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →