Atlas

Models

← All models

Claude Mythos Preview

Anthropic · 2026-04-07

Claude Mythos Preview is a general-purpose frontier model with advanced agentic coding and reasoning skills and a major increase in cybersecurity capability over prior Claude models. Anthropic made it available only to invited Project Glasswing partners for defensive cybersecurity work; on the API, adaptive thinking is the default and thinking depth can be varied with the effort parameter.

Benchmark scores

BenchmarkScore
BrowseComp87.9
CAISI PortBench80.1
Capability166.3
CharXiv Reasoning86.1
CharXiv Reasoning (with tools)93.2
CTF-Archive Diamond65.6
Cybench100.0
CyberGym83.1
DeepSearchQA (F1 score)94.4
ExploitBench (CAISI Inspect + AutoNudge)57.2
ExploitBench Leaderboard (Capability coverage)78.0
ExploitBench Leaderboard (Mean capability score, out of 16)10.0
ExploitGym18.1
ExploitGym intended-vulnerability exploits with mitigations45.0
ExploitGym kernel exploits with mitigations3.0
ExploitGym kernel intended-vulnerability exploits12.0
ExploitGym userspace exploits with mitigations25.0
ExploitGym userspace intended-vulnerability exploits107
ExploitGym V8 exploits with mitigations17.0
ExploitGym V8 intended-vulnerability exploits38.0
FrontierScience Olympiad83.0
GPQA Diamond92.9
GraphWalks BFS - 1M74.3
GraphWalks BFS - 256K85.7
Humanity's Last Exam64.7
METR Time Horizons1045
MMMLU92.7
OSWorld79.6
OSWorld-Verified85.4
OTIS-AIME 2025 (CAISI 30-task set)99.5
Loading Atlas data…