GBA Eval: leaderboard
A coding agent gets 24 hours in a container to write a Game Boy Advance emulator in Rust from scratch, compiled to WebAssembly, given a hardware reference, example ROMs and a black box reference emulator. It is graded against Mesen2 on gameplay replays, self checking test ROMs and audio. The overall score, from 0 to 1, weights replays 60 percent and procedural and audio tests 20 percent each; higher is better.
Metric: Overall Emulator Score. Source: github.com. Source: gbaeval.com. Saturation forecast: Around December 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 0.8 | #12 |
| 2 | Claude Fable 5 | 0.74 | #16 |
| 3 | Claude Opus 4.8 | 0.71 | #37 |
| 4 | Grok 4.5 | 0.65 | #48 |
| 5 | Claude Sonnet 5 | 0.65 | #57 |
| 6 | GPT-5.5 | 0.53 | #31 |
| 7 | GPT-5.6 Sol | 0.53 | #19 |
| 8 | Claude Sonnet 4.6 | 0.49 | #90 |
| 9 | Kimi K3 | 0.48 | #30 |
| 10 | GPT-5.4 | 0.45 | #79 |
| 11 | Claude Opus 4.6 | 0.44 | #63 |
| 12 | Claude Opus 4.7 | 0.44 | #49 |
| 13 | Muse Spark 1.1 | 0.08 | #52 |
| 14 | Gemini 3.5 Flash | 0.07 | #60 |
| 15 | Grok Build 0.1 | 0.02 | #117 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gba-eval · How It Works · Data refreshed daily, snapshot 2026-10-20.