GBAEval

What does it measure?
Made by Mechanize: tests whether an AI can build, from scratch, a program that imitates the Game Boy Advance console (an emulator).
Who evaluates it, and how?
A single task: the AI builds it in Rust, offline, over 24 hours. It is graded by comparing screens, behavior, and audio with a reference emulator (Mesen2); replay tests carry the most weight.
Reading the score
Scores run from 0 to 100%. Higher is better.
Top score90.6%
Models tested28
Scoring basisHigher is better
Last updated2026.11.14
Model release date
  1. 1
    Anthropic2026.09.22
    90.6%
  2. 2
    OpenAI2026.09.04
    87.1%
  3. 3
    Anthropic2026.11.09
    84.8%
  4. 4
    OpenAI2026.09.29
    81.9%
  5. 5
    Anthropic2026.07.24
    79.6%
  6. 6
    Google2026.09.30
    78.7%
  7. 7
    Anthropic2026.06.09
    74.5%
  8. 8
    Anthropic2026.05.27
    70.9%
  9. 9
    SpaceXSI2026.07.08
    65.4%
  10. 10
    Anthropic2026.06.30
    65.3%