Model Drop

Claude Mythos Preview

April 7, 2026 2 min read

Claude Mythos Preview, which Anthropic announced on April 7, 2026, is a general-purpose frontier model so good at finding and exploiting software flaws that Anthropic gave it only to Project Glasswing security partners instead of releasing it.

Model: Claude Mythos Preview (claude-mythos-preview)
Maker: Anthropic
Ship date: April 7, 2026 (invite-only research preview)
Pricing: $25 / $125 per M tokens (Glasswing participants only, after $100M in usage credits)
Context: 1M tokens
Where you can use it:
- Project Glasswing partners only, by invite
- Claude API, Bedrock, Vertex AI, Foundry
When to reach for it:
- Finding and fixing security flaws
- Hardening critical open-source code
- Skip it unless you're a Glasswing partner
How it compares:
- Far ahead of GPT-5.4 on hard coding tasks
- Close to Gemini 3.1 Pro on science Q&A
What makes it different:
- Found thousands of zero-day vulnerabilities
- Withheld from public release over cyber risk
Early reactions: Even observers wary of "too dangerous to release" hype called the caution warranted, as maintainers reported a flood of real AI-found bugs.
Read more: Project Glasswing, Claude Mythos Preview system card, Assessing Claude Mythos Preview's cybersecurity capabilities

The benchmarks

Anthropic ran Mythos Preview with adaptive thinking at max effort, averaged over 5 trials, against Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro using rival numbers from their makers; OpenAI's Terminal-Bench 2.0 score used its own harness, Anthropic printed Gemini 3.1 Pro's MMMLU as a 92.6% to 93.6% range, so it's left out of that row, and memorization screens flagged some SWE-bench problems, though Anthropic says the lead holds without them (it also says strong low-effort Humanity's Last Exam scores could mean some memorization). It leads every row in Anthropic's table, so no row below shows it trailing. Nobody outside the preview could test it, so independent numbers are scarce.

Anthropic's numbers

SWE-bench Verified (real GitHub fixes): 93.9%. +13.1 over Claude Opus 4.6 (80.8%), +13.3 over Gemini 3.1 Pro (80.6%).
SWE-bench Pro (harder real-repo fixes): 77.8%. +20.1 over GPT-5.4 (57.7%), +23.6 over Gemini 3.1 Pro (54.2%), +24.4 over Claude Opus 4.6 (53.4%).
Terminal-Bench 2.0 (command-line agent tasks): 82.0%. +6.9 over GPT-5.4 (75.1%), +13.5 over Gemini 3.1 Pro (68.5%), +16.6 over Claude Opus 4.6 (65.4%).
CyberGym (reproducing real vulnerabilities): 83.1%. +16.5 over Claude Opus 4.6 (66.6%).
GPQA Diamond (grad-level science): 94.5%. +0.2 over Gemini 3.1 Pro (94.3%), +1.7 over GPT-5.4 (92.8%), +3.2 over Claude Opus 4.6 (91.3%).
Humanity's Last Exam (no tools) (expert questions): 56.8%. +12.4 over Gemini 3.1 Pro (44.4%), +17.0 over GPT-5.4 (39.8%), +16.8 over Claude Opus 4.6 (40.0%).
OSWorld-Verified (operating a computer): 79.6%. +4.6 over GPT-5.4 (75.0%), +6.9 over Claude Opus 4.6 (72.7%).
MMMLU (multilingual knowledge): 92.7%. +1.6 over Claude Opus 4.6 (91.1%).

Independent numbers

UK AISI expert-level capture-the-flag (cyber challenges): 73%.
UK AISI "The Last Ones" cyber range (32-step network attack): 22 of 32 steps. +6 over Claude Opus 4.6 (16 steps).