Skip to content

Benchmark

MLE-Bench Lite

Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages. Higher is better.

Models measured
18
Scores recorded
33
Sources
3
IQ weight
2.65

Score distribution

59%
Mean
66.6%
Median
22%
Std dev
18.6%
10th percentile
78.8%
90th percentile
4.6%
Min
4.6%80.3%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 2.65 for chat models, measured over 18 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Gemini 3 Pro PreviewFamou-Agent 2.0· MLE-bench official data80.3%
2Gemini 3 Pro PreviewMLEvolve· MLE-bench official data80.3%
3Gemini 3 Pro PreviewPiEvolve (Fractal AI Research)· MLE-bench official data80.3%
4Gemini 3 Pro PreviewCAIR MARS+· MLE-bench official data78.8%
5Claude Opus 4.6AIBuildAI· MLE-bench official data77.3%
6Gemini 3 FlashLoongFlow· MLE-bench official data77.3%
7DeepSeek V3.2 SpecialeML-Master 2.0· MLE-bench official data75.8%
8Gemini 2.5 ProFamou-Agent 2.0· MLE-bench official data75.8%
9Claude Opus 4.6MiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.75.7%
10Gemini 3 Pro PreviewCAIR MARS· MLE-bench official data74.2%
11Claude Sonnet 4.6MiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.72.7%
12GPT 5.4MiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.71.2%
13Gemini 2.5 ProCAIR MLE-STAR-Pro-1.5· MLE-bench official data68.2%
14Gemini 3 Pro PreviewLeeroo· MLE-bench official data68.2%
15GPT 5R&D-Agent· MLE-bench official data68.2%
16Gemini 2.5 ProCAIR MLE-STAR-Pro-1.0· MLE-bench official data66.7%
17MiniMax M2.7· LLM Stats data· Reported by the model's developer. The source relaying it did not verify it. claim66.6%
18Gemini 3.1 Pro PreviewMiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.66.6%
19MiniMax M2.7MiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.66.6%
20GPT-5 CodexThesis· MLE-bench official data65.2%
21GPT 5Operand ensemble· MLE-bench official data63.6%
22DeepSeek R1InternAgent· MLE-bench official data62.1%
23Gemini 2.5 ProFamou-Agent· MLE-bench official data62.1%
24OpenAI o3AIRA-dojo· MLE-bench official data55%
25MiniMax M2.5MiniMax M2.7 official benchmark overview · exact harness/version not stated· Reported by MiniMax· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it.51.5%
26DeepSeek R1ML-Master· MLE-bench official data48.5%
27O1 PreviewR&D-Agent· MLE-bench official data48.2%
28O1 PreviewAIDE· MLE-bench official data35.9%
29Claude 3.5 SonnetAIDE· MLE-bench official data19.7%
30GPT 4oAIDE· MLE-bench official data18.6%
31GPT 4oOpenHands· MLE-bench official data12.1%
32Llama 3.1 405B InstructAIDE· MLE-bench official data10.2%
33GPT 4oMLAB· MLE-bench official data4.6%

Data credits: 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. · MLE-bench (OpenAI), MIT (licence) · Data by LLM Stats