SYLIN

Claude Mythos

by Anthropic · Model · Language & code models

promising

Mythos 5.2 looks stronger than 5.1 on security tasks, per vendor-side numbers only

Devin added Mythos 5.2 to security scans, citing 73.5 versus 56.9 for Mythos 5.1 on an internal cyber benchmark

Why we say so

Cognition, which added it to Devin's security scans, cites 73.5 versus 56.9 for 5.1 on its own internal cyber benchmark, so this is a company's own test, not independent. Anthropic's system card, as summarized by dbreunig, shows a small negative self-preference bias when grading Claude transcripts (0.10 lower on a 0-9 scale). No independent benchmarks or hands-on reviews yet.

Versions

VersionReleasedRank
Claude Mythos 5.2LATEST
promising Big internal-benchmark gain over 5.1, not yet independently confirmed
Nov 11
Claude Mythos 5.1
unproven Only appears as the comparison baseline for 5.2
Sep 1

Evidence

  1. Nov 11
    System card: Mythos 5.2 shows small negative self-preference bias, grading Claude transcripts 0.10 lower on a 0-9 scale
  2. Sep 3
    Mythos 5.1, Fable without guardrails and not public, designed lab-tested protein binders binding 10x stronger; nearly 50% of designs worked
  3. Sep 1
    Mythos 5.1 released only through trusted-access programs for cybersecurity and life sciences

Alternatives

KimiCost-versus-performance charts place K3 as more cost-efficient per task than Claude Mythos and Opus
AISLEAISLE claims it found curl CVEs where Anthropic's Mythos and OpenAI's Codex reported zero
GLMClaimed best on CyberGym, ahead of Mythos and GPT 5.6 Sol, and well ahead of Kimi K3 on ExploitBench and ExploitGym