Benchmark
MLE-Bench
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.
- Models measured
- 20
- Scores recorded
- 34
- Sources
- 3
- IQ weight
- 1.16
Score distribution
- 42%
- Mean
- 44%
- Median
- 19.7%
- Std dev
- 7.6%
- 10th percentile
- 63.9%
- 90th percentile
- 1.6%
- Min
This benchmark has a weight of 1.16 for chat models, measured over 20 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | Score |
|---|---|---|
| 1 | Claude Sonnet 5Gemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 66.9% |
| 2 | Gemini 3 Pro PreviewFamou-Agent 2.0· MLE-bench official data | 64.4% |
| 3 | Gemini 3.6 Flash· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 63.9% |
| 4 | Gemini 3.6 FlashGemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 63.9% |
| 5 | Claude Opus 4.6AIBuildAI· MLE-bench official data | 63.1% |
| 6 | Gemini 3 Pro PreviewCAIR MARS+· MLE-bench official data | 62.7% |
| 7 | Gemini 3 FlashLoongFlow· MLE-bench official data | 62.7% |
| 8 | Gemini 3 Pro PreviewMLEvolve· MLE-bench official data | 61.3% |
| 9 | Gemini 3 Pro PreviewPiEvolve (Fractal AI Research)· MLE-bench official data | 61.3% |
| 10 | Gemini 2.5 ProFamou-Agent 2.0· MLE-bench official data | 59.6% |
| 11 | DeepSeek V3.2 SpecialeML-Master 2.0· MLE-bench official data | 56.4% |
| 12 | Gemini 3 Pro PreviewCAIR MARS· MLE-bench official data | 56% |
| 13 | Gemini 3 Pro PreviewLeeroo· MLE-bench official data | 50.7% |
| 14 | Gemini 3.5 FlashGemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 49.7% |
| 15 | GPT-5 CodexThesis· MLE-bench official data | 48.4% |
| 16 | GPT-5.6 LunaGemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 47.6% |
| 17 | Gemini 2.5 ProCAIR MLE-STAR-Pro-1.5· MLE-bench official data | 44% |
| 18 | Gemini 2.5 ProFamou-Agent· MLE-bench official data | 43.6% |
| 19 | Grok 4.5Gemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 43.2% |
| 20 | Gemini 3.1 Pro PreviewGemini 3.6 Flash Model Card · July 2026· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 42.6% |
| 21 | GPT 5Operand ensemble· MLE-bench official data | 39.6% |
| 22 | Gemini 3.5 Flash Lite· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim | 39.2% |
| 23 | Gemini 2.5 ProCAIR MLE-STAR-Pro-1.0· MLE-bench official data | 38.7% |
| 24 | DeepSeek R1InternAgent· MLE-bench official data | 36.4% |
| 25 | GPT 5R&D-Agent· MLE-bench official data | 35.1% |
| 26 | OpenAI o3AIRA-dojo· MLE-bench official data | 31.6% |
| 27 | DeepSeek R1ML-Master· MLE-bench official data | 29.3% |
| 28 | O1 PreviewR&D-Agent· MLE-bench official data | 22.4% |
| 29 | O1 PreviewAIDE· MLE-bench official data | 17.1% |
| 30 | GPT 4oAIDE· MLE-bench official data | 8.6% |
| 31 | Claude 3.5 SonnetAIDE· MLE-bench official data | 7.6% |
| 32 | GPT 4oOpenHands· MLE-bench official data | 4.9% |
| 33 | Llama 3.1 405B InstructAIDE· MLE-bench official data | 3.3% |
| 34 | GPT 4oMLAB· MLE-bench official data | 1.6% |
Data credits: MLE-bench (OpenAI), MIT (licence) · BenchBoard by SYuan03 (github.com/SYuan03/benchboard). MIT covers its code and data organization; no results licence is inferred. Numeric relays retain the original stating-source links. · Data by LLM Stats