Benchmark
hal-online-mind2web
Agentic benchmarks measure how well models carry out tasks that take several steps. Scores are percentages.
All agentic benchmarksScores from HAL (Princeton).
- Models measured
- 9
- Scores recorded
- 12
- Sources
- 1
- IQ weight
- 0.89
Score distribution
- 35.4%
- Mean
- 38.3%
- Median
- 5%
- Std dev
- 29%
- 10th percentile
- 40%
- 90th percentile
- 25.3%
- Min
This benchmark has a weight of 0.89 for chat models, measured over 9 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 | GPT 5medium· HAL (Princeton) | 42.3% |
| 2 | Claude Sonnet 4· HAL (Princeton) | 40% |
| 3 | Claude Sonnet 3.7high· HAL (Princeton) | 39.3% |
| 4 | Claude Sonnet 4high· HAL (Princeton) | 39.3% |
| 5 | OpenAI o3medium· HAL (Princeton) | 39% |
| 6 | Claude Sonnet 3.7· HAL (Princeton) | 38.3% |
| 7 | GPT 4.1· HAL (Princeton) | 36.3% |
| 8 | DeepSeek V3· HAL (Princeton) | 32.3% |
| 9 | o4 Minihigh· HAL (Princeton) | 32% |
| 10 | o4 Minilow· HAL (Princeton) | 31.7% |
| 11 | Gemini 2.0 Flash· HAL (Princeton) | 29% |
| 12 | DeepSeek R1· HAL (Princeton) | 25.3% |
Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.
Data credits: HAL (Princeton) — cite the HAL paper