Epoch AI - Scicode: leaderboard
SciCode turns 80 research coding problems from the natural sciences into subproblems, each solved by a Python function that must pass the benchmark's test cases. Epoch AI does not run it; this board republishes the subproblem accuracy measured by Artificial Analysis, the share of subproblems whose code passes, where higher is better. Source: artificialanalysis.ai/evaluations/scicode.
Metric: Score. Source: epoch.ai. Saturation forecast: Around 2028. 221 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 4 Argon (High) | 61.81 | #8 (Gemini 4 Argon) |
| 2 | MiMo-V2.6-Pro | 60.88 | #42 |
| 3 | Muse Wonder (Max) | 60.3 | #15 (Muse Wonder) |
| 4 | Gemini 3.7 Flash (Medium) | 59.84 | #23 (Gemini 3.7 Flash) |
| 5 | Muse Spark 1.3 (xHigh) | 59.72 | #24 (Muse Spark 1.3) |
| 6 | Kimi K3 (Max) | 59.49 | #30 (Kimi K3) |
| 7 | GLM-5.3 (Max) | 59.03 | #39 (GLM-5.3) |
| 8 | Step 5 Preview | 58.91 | #44 |
| 9 | Muse Spark 1.3 (Max) | 58.8 | #24 (Muse Spark 1.3) |
| 10 | Muse Spark 1.1 (xHigh) | 58.8 | #52 (Muse Spark 1.1) |
| 11 | Gemini 3.1 Pro (Preview) | 58.68 | #58 |
| 12 | Kimi K3 | 58.68 | #30 |
| 13 | Muse Spark Open (xHigh) | 58.45 | #54 (Muse Spark Open) |
| 14 | GPT-5.6 Sol (High) | 57.75 | #19 (GPT-5.6 Sol) |
| 15 | Grok 4.7 (High) | 57.75 | #25 (Grok 4.7) |
No result here: #3 Claude Opus 5.5, #9 Claude Fable 5.1, #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=epoch-ai-scicode · How It Works · Data refreshed daily, snapshot 2026-10-20.