Claude Mythos Preview

What it is

Claude Mythos Preview is an unreleased general-purpose frontier model, announced by Anthropic on 7 April 2026 together with Project Glasswing, a restricted programme for securing critical software. A frontier model is one of the most capable systems at the leading edge of AI development.^[1]

Origins

Anthropic said that the model's coding ability let it surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Using it, the company said it had found thousands of high-severity vulnerabilities, including in every major operating system and web browser.^[1]

Some examples given were a 27-year-old flaw in OpenBSD that let an attacker remotely crash any machine running it, a 16-year-old bug in FFmpeg, and chained Linux kernel vulnerabilities that led to full control of a machine. Anthropic said all of them were reported and patched.^[1]

Mozilla reported that it had found and patched 271 vulnerabilities in Firefox. According to Wikipedia, access was expanded to 150 cybersecurity organisations on 2 June 2026.^[4]

Main features

Anthropic compared the model with Claude Opus 4.6. On SWE-bench Verified it scored 93.9% against 80.8%. On SWE-bench Pro it scored 77.8% against 53.4%, on Terminal-Bench 2.0 82.0% against 65.4%, on CyberGym 83.1% against 66.6%, on GPQA Diamond 94.6% against 91.3%, and on OSWorld-Verified 79.6% against 72.7%.^[1]

On Humanity's Last Exam, a hard academic test, it scored 56.8% without tools and 64.7% with tools.^[1]

Anthropic also noted that the model uses its newer tokenizer, the component that splits text into tokens, and includes the full 1M-token context window at standard pricing.^[2]

Access and pricing

Anthropic did not plan general availability. It cited the risk that the capabilities could spread to actors who would not deploy them safely, and said it first needed safeguards to block the most dangerous outputs.^[1]

Launch partners were Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, along with more than 40 other organisations that maintain critical software.^[1]

Anthropic committed up to $100M in usage credits and $4M in donations to open-source security groups. After the credits, the price was $25 per million input tokens and $125 per million output tokens, through the Claude API, Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry.^[1]

Its place in development

Anthropic called the model its best-aligned model when it released Claude Opus 4.7.^[5] Claude Mythos 5 succeeded Mythos Preview on 9 June 2026, when Mythos Preview was deprecated.^[3][6]