Anthropic releases a version of its vaunted Mythos model to developers

Anthropic released Claude Fable 5 on Tuesday, a publicly accessible version of its advanced "Mythos" AI model. The company stated that Claude Fable 5 surpasses all previous public releases in capabilities, demonstrating exceptional performance in software engineering, knowledge work, and vision tasks, and is superior in handling extended, intricate tasks. Hex, an analytics firm, reported that Fable 5 was the first model to achieve over 90% on its benchmark for complex, long-duration analytical tasks, marking a 10-point improvement over earlier Opus models. In coding agentic skills measured by SWE-Bench-Pro, Claude Fable 5 scored 80.3%, significantly outperforming OpenAI's GPT-5.5 (58.6%) and Google Deepmind's Gemini 3.1 Pro (54.2%), with testers noting its "strong judgment and attention to nuance."

Anthropic initially released the Mythos Preview model in April to a limited group of cybersecurity professionals and critical infrastructure overseers, who found it effective at identifying and exploiting software security vulnerabilities. The new Fable 5 model includes specific restrictions, such as deferring cybersecurity and biology-related queries to the less advanced Claude Opus 4.8 model, which occurred in approximately five percent of sessions. Anthropic confirmed that internal and external teams attempted to elicit responses on prohibited topics, but all attempts were unsuccessful.

Cybersecurity researchers who previously had access to Mythos Preview will now receive Claude Mythos 5. Anthropic plans to expand access through a "more systematic trusted-access program." This public release of a Mythos-derived model follows closely on the heels of other developments from the company.