AlgoTune

What does it measure?
Tests rewriting code for computational tasks such as compression, encryption, matrix decomposition and optimization so it runs faster than the existing implementation.
Who evaluates it, and how?
154 tasks collected from domain experts, timed against reference implementations from open-source packages. The result must stay correct, and the score is the average number of times faster than the reference.
Reading the score
Scores are the speed-up over the baseline. Higher is better.
Top score2.02×
Models tested11
Scoring basisHigher is better
Model release date
  1. 1
    Google2026.02.19
    2.02×
  2. 2
    OpenAI2026.03.05 · high
    1.85×
  3. 3
    Anthropic2025.11.24
    1.77×
  4. 4
    OpenAI2025.08.07 · high
    1.67×
  5. 5
    Anthropic2025.09.29
    1.52×
  6. 6
    Google2025.06.17
    1.51×
  7. 7
    Anthropic2026.02.04 · high
    1.47×
  8. 8
    OpenAI2025.08.05 · high
    1.41×
  9. 9
    OpenAI2025.08.07 · high
    1.38×
  10. 10
    Anthropic2025.08.05
    1.34×

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.