Benchmark
hal-gaia
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages.
All agentic benchmarksScores from HAL (Princeton).
- Models measured
- 12
- Scores recorded
- 17
- Sources
- 1
- IQ weight
- 1.27
Score distribution
- 54.7%
- Mean
- 57.6%
- Median
- 14.2%
- Std dev
- 32.7%
- 10th percentile
- 68.5%
- 90th percentile
- 29.4%
- Min
This benchmark has a weight of 1.27 for chat models, measured over 12 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model | |
|---|---|---|
| 1 | Claude Sonnet 4.5· HAL (Princeton) | 74.5% |
| 2 | Claude Sonnet 4.5high· HAL (Princeton) | 70.9% |
| 3 | Claude Opus 4.1high· HAL (Princeton) | 68.5% |
| 4 | Claude Opus 4high· HAL (Princeton) | 64.9% |
| 5 | Claude Opus 4.1· HAL (Princeton) | 64.2% |
| 6 | Claude Sonnet 3.7high· HAL (Princeton) | 64.2% |
| 7 | GPT 5medium· HAL (Princeton) | 62.8% |
| 8 | o4 Minilow· HAL (Princeton) | 58.2% |
| 9 | Claude Opus 4· HAL (Princeton) | 57.6% |
| 10 | Claude Haiku 4.5· HAL (Princeton) | 56.4% |
| 11 | Claude Sonnet 3.7· HAL (Princeton) | 56.4% |
| 12 | o4 Minihigh· HAL (Princeton) | 55.8% |
| 13 | GPT 4.1· HAL (Princeton) | 50.3% |
| 14 | Gemini 2.0 Flash· HAL (Princeton) | 32.7% |
| 15 | OpenAI o3medium· HAL (Princeton) | 32.7% |
| 16 | DeepSeek R1· HAL (Princeton) | 30.3% |
| 17 | DeepSeek V3· HAL (Princeton) | 29.4% |
Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.
Data credits: HAL (Princeton) — cite the HAL paper