EnigmaEval

What does it measure?
Made by Scale AI and partners: long puzzle-hunt puzzles that take creative thinking and many chained deductions.
Who evaluates it, and how?
1,184 puzzles mixing text and images. Only the final answer is checked, and it must match the key exactly; intermediate steps are not graded.
Reading the score
Scores run from 0 to 100%. Higher is better. Puzzles come from beginner-friendly hunts and from hard events like the MIT Mystery Hunt; on the hard set even frontier models score near zero.
Top score39.3%
Models tested19
Scoring basisHigher is better
Model release date
  1. 1
    Anthropic2026.06.09 · high
    39.3%
  2. 2
    OpenAI2026.07.09 · high
    37.1%
  3. 3
    Google2026.02.19
    36.8%
  4. 4
    Google2026.05.19 · high
    25.4%
  5. 5
    OpenAI2026.03.05
    23.8%
  6. 6
    Anthropic2026.05.27 · xhigh
    23.5%
  7. 7
    OpenAI2026.03.05 · xhigh
    16.0%
  8. 8
    Anthropic2025.11.24
    11.9%
  9. 9
    OpenAI2025.08.07
    10.5%
  10. 10
    OpenAI2025.08.07
    8.2%

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.