MLS-Bench Lite: leaderboard
Official 30-task subset of MLS-Bench for evaluating whether AI systems can invent generalizable and scalable machine-learning methods under a five-hour agent budget. Preserves each published harness, reasoning-effort and fallback configuration.
Metric: Score. Source: mls-bench.com. Source: mls-bench.com. Saturation forecast: Rough model projection: around 2026. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (max) (MLS-Bench Run: model%3DClaude%20Fable%205.1%3Bharness%3DClaude%20Code%20max) | 50.3 | |
| 2 | GPT-6 Astra (max) (MLS-Bench Run: model%3DGPT-6%20Astra%3Bharness%3DCodex%20max) | 50 | |
| 3 | Claude Fable 5 (max) (MLS-Bench Run: model%3DClaude%20Fable%205%3Bharness%3DClaude%20Code%20max%20%28%20with%20fallback%20%29) | 49.9 | |
| 4 | Claude Opus 5 (max) (MLS-Bench Run: model%3DClaude%20Opus%205%3Bharness%3DClaude%20Code%20max) | 49.8 | |
| 5 | Kimi K3 (max) (MLS-Bench Run: model%3DKimi%20K3%3Bharness%3DKimi-Code%20max) | 48.3 | |
| 6 | GPT 5.6 Sol (max) (MLS-Bench Run: model%3DGPT%205.6%20Sol%3Bharness%3DCodex%20max) | 46.2 | |
| 7 | Claude Opus 4.8 (max) (MLS-Bench Run: model%3DClaude%20Opus%204.8%3Bharness%3DClaude%20Code%20max) | 42.8 | |
| 8 | Qwen3.8-Max (MLS-Bench Run: model%3DQwen3.8-Max%3Bharness%3DClaude%20Code) | 41 | |
| 9 | GLM 5.2 (max) (MLS-Bench Run: model%3DGLM%205.2%3Bharness%3DClaude%20Code%20max) | 40.4 | |
| 10 | GPT-5.5 (xhigh) (MLS-Bench Run: model%3DGPT-5.5%3Bharness%3DCodex%20xhigh) | 35.5 |
Interactive version: theaggregate.ai/benchmark?slug=mls-bench-lite · How It Works · Data refreshed daily, snapshot 2026-10-20.