Claude Mythos Preview

2026-04-07 · Publisher : Anthropic · Domain : Artificial Intelligence

Anthropic announces Claude Mythos Preview, a frontier cybersecurity model reserved for Project Glasswing partners, able to find and exploit software vulnerabilities, with no public release.

Anthropic announces Claude Mythos Preview, a frontier cybersecurity model reserved for Project Glasswing partners, able to find and exploit software vulnerabilities, with no public availability. ### In detail - Restricted, invitation-only research preview for cyber defense (Project Glasswing). - According to the vendor, the model discovered thousands of zero-day vulnerabilities in major operating systems and browsers. - Partners named: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and more than 40 organizations. - No general availability planned. Deprecated on June 9, 2026 with the arrival of Mythos 5. ### Specifications - Access: Claude API, Amazon Bedrock, Vertex AI, Microsoft Foundry, by invitation only. API identifier: `claude-mythos-preview`. - 1M-token context according to the pricing page; maximum output and knowledge cutoff: not disclosed. - New tokenizer (about 30% more tokens for the same text). ### Pricing | Item | Pricing (USD per million tokens) | | --- | --- | | Input (after an initial credit of 100 million USD in usage offered by Anthropic) | 25 | | Output | 125 | ### Benchmarks Scores published by the vendor (comparison with Opus 4.6). | Evaluation | Mythos Preview | Opus 4.6 | | --- | --- | --- | | CyberGym | 83.1% | 66.6% | | SWE-bench Verified | 93.9% | 80.8% | | Terminal-Bench 2.0 | 82.0% | 65.4% | | GPQA Diamond | 94.6% | 91.3% | ### Key takeaways The first Anthropic model deliberately withheld for its offensive capabilities, it opens the restricted-access approach later taken up by the Mythos 5 and 5.1 models.

Tags : LLM, Anthropic, Claude, Mythos, Cybersécurité, Accès restreint