SnakeBench

Citation frequency 1/3
What does it measure?
A fun benchmark where two AIs play the snake game head to head on the same board to see which one survives longer.
Who evaluates it, and how?
Each turn the model gets the board state (snake and apple positions) as text and picks the next direction. A snake that hits a wall or a snake body is out, and the accumulated results are turned into a TrueSkill rating (a score from head-to-head results). We list only models whose uncertainty (sigma) is below 3, the level the site calls confident.
Reading the score
Scores are ratings converted from contest results. Higher is better. Run by Greg Kamradt as a side project for the ARC Prize Foundation. Games keep accumulating, so ratings change often, and most models have only about 10 to 35 games. It is a score for fun and is not used in our use-case rankings.
Top score39
Models tested300
Scoring basisHigher is better
Last updated2026.11.21
Model release date
240 models without a release date are not plotted; they appear only in the table below.
  1. 1
    OpenAI2026.04.23
    39
  2. 2
    Google2026.02.19
    38
  3. 3
    Google
    gemini-3-pro-preview
    38
  4. 4
    OpenAI2026.11.17
    38
  5. 5
    OpenAI2026.09.04
    37
  6. 6
    OpenAI
    gpt-5.1-codex-mini
    36
  7. 7
    o3
    36
  8. 8
    Anthropic2026.05.27
    35
  9. 9
    gemini-pro-latest
    35
  10. 10
    OpenAI2025.08.07
    35