Claude Mythos Preview
Claude Mythos Preview, a tier above Opus, out-finds all but the most skilled humans at software vulnerabilities, so Anthropic gates it to about 50 organizations.
Anthropic's strongest model to date, withheld from general release for cyber risk. Versus Opus 4.6: SWE-bench Verified 93.9% vs 80.8%, SWE-bench Pro 77.8% vs 53.4%, Terminal-Bench 2.0 82.0% vs 65.4%, CyberGym 83.1% vs 66.6%. Participants pay $25/$125 per million tokens after credits.
- Date
- Tuesday, 7 April 2026
- Lab
- Anthropic
- Kind
- model
- Access
- research preview
- Price
- $25/$125 per M tokens in/out for Glasswing participants after credits (2026-04-07)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 93.9% vs 80.8% for Opus 4.6 | company |
| SWE-bench Pro | 77.8% vs 53.4% for Opus 4.6 | company |
| Terminal-Bench 2.0 | 82.0% vs 65.4% for Opus 4.6 | company |
| CyberGym vulnerability reproduction | 83.1% vs 66.6% for Opus 4.6 | company |
Gated release with no plan for general availability. Access is via the Claude API, Bedrock, Vertex AI and Foundry for vetted organizations. About 50 initial partners per Anthropic's later Glasswing updates, and the successor Claude Mythos 5 shipped 2026-06-09. Anthropic's Opus 4.7 page calls Mythos Preview its best-aligned model (company claim). OpenAI staff said GPT-5.5 would bolster its cyber deployment in response to Mythos (TechCrunch). Anthropic's own test found that Zhipu's open-weights GLM-5.3 reached full control-flow hijacks in 4% of trials vs 6% for Mythos Preview (see anthropic-glm-5-3-cyber-spread).
Sources
- www.anthropic.com/glasswing
- www.anthropic.com/research/mythos-preview
- www.anthropic.com/news/claude-opus-4-7
- techcrunch.com/2026/04/23/openai-chatgpt-gpt-5-5-ai-model-superapp/
This record was checked against its sources on 6 October 2026. How we check