Claude Mythos Preview
FrontierAnthropic•Released on 2026-04-07
Anthropic's most powerful model ever, released April 7, 2026 through Project Glasswing. Not publicly available due to cybersecurity capabilities. Achieved breakthrough performance: 93.9% SWE-bench Verified (+13pp over Opus 4.6), autonomously discovered zero-day vulnerabilities in every major OS and browser. Internal codename was 'Capybara'. Anthropic describes it as a 'step change in capabilities' and 'distinct model category' above Opus tier.
Core Specs
0K
Context Window
0K
Max Output
ReasoningOpen Sourcetextimage
Pros & Cons
Pros
- +Highest SWE-bench Verified ever (93.9%)
- +Breakthrough cybersecurity capabilities - finds zero-days in major OS/browsers
- +77.8% SWE-bench Pro (20pp above GPT-5.4)
- +59% SWE-bench Multimodal (2x Opus 4.6)
- +Strong agentic coding and reasoning
Cons
- −Not publicly available - Project Glasswing partners only
- −No API access for general users
- −40 organizations total have access
- −Dual-use risk: same capabilities that find vulnerabilities can exploit them
- −Timeline for general availability unknown
Pricing
Input (per 1M tokens)$0.00
Output (per 1M tokens)$0.00
Updated on 2026-04-08
Get Started
1Visit the provider's website
2Create an account
3Start using the model
Benchmarks
sweBenchVerified93.9%
sweBenchPro77.8%
sweBenchMultilingual87.3%
sweBenchMultimodal59%
terminalBench282%
usamo202697.6%
noteAll benchmarks represent significant improvements over Opus 4.6 and GPT-5.4%
Reliability
Incidents (30d)0
Only available to 40 organizations via Project Glasswing