Benchmark
hal-scienceagentbench
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages.
All agentic benchmarksScores from HAL (Princeton).
- Models measured
- 11
- Scores recorded
- 16
- Sources
- 1
- IQ weight
- 1.99
Score distribution
- 25.2%
- Mean
- 27.4%
- Median
- 5.5%
- Std dev
- 18.6%
- 10th percentile
- 30.4%
- 90th percentile
- 12.8%
- Min
This benchmark has a weight of 1.99 for chat models, measured over 11 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 | OpenAI o3medium· HAL (Princeton) | 33.3% |
| 2 | Claude Sonnet 3.7high· HAL (Princeton) | 30.4% |
| 3 | Claude Sonnet 4.5high· HAL (Princeton) | 30.4% |
| 4 | GPT 5medium· HAL (Princeton) | 30.4% |
| 5 | Claude Sonnet 4.5· HAL (Princeton) | 29.4% |
| 6 | Claude Opus 4.1· HAL (Princeton) | 27.4% |
| 7 | o4 Minihigh· HAL (Princeton) | 27.4% |
| 8 | o4 Minilow· HAL (Princeton) | 27.4% |
| 9 | Claude Opus 4.1high· HAL (Princeton) | 26.5% |
| 10 | GPT 4.1· HAL (Princeton) | 24.5% |
| 11 | Claude Haiku 4.5high· HAL (Princeton) | 23.5% |
| 12 | DeepSeek R1· HAL (Princeton) | 23.5% |
| 13 | Claude Sonnet 3.7· HAL (Princeton) | 22.6% |
| 14 | Claude Haiku 4.5· HAL (Princeton) | 18.6% |
| 15 | DeepSeek V3· HAL (Princeton) | 15.7% |
| 16 | Gemini 2.0 Flash· HAL (Princeton) | 12.8% |
Data credits: HAL (Princeton) — cite the HAL paper