Benchmark
hal-usaco
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
- 1.25
Score distribution
- 44.8%
- Mean
- 46.3%
- Median
- 13.2%
- Std dev
- 27%
- 10th percentile
- 64.8%
- 90th percentile
- 26.7%
- Min
This benchmark has a weight of 1.25 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) | 69.1% |
| 2 | o4 Minihigh· HAL (Princeton) | 64.8% |
| 3 | o4 Minilow· HAL (Princeton) | 53.1% |
| 4 | Claude Opus 4.1high· HAL (Princeton) | 51.5% |
| 5 | Claude Opus 4.1· HAL (Princeton) | 48.2% |
| 6 | OpenAI o3medium· HAL (Princeton) | 46.3% |
| 7 | GPT 4.1· HAL (Princeton) | 45% |
| 8 | DeepSeek V3· HAL (Princeton) | 39.1% |
| 9 | DeepSeek R1· HAL (Princeton) | 38.1% |
| 10 | Claude Sonnet 3.7· HAL (Princeton) | 29.3% |
| 11 | Gemini 2.0 Flash· HAL (Princeton) | 27% |
| 12 | Claude Sonnet 3.7high· HAL (Princeton) | 26.7% |
Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.
Data credits: HAL (Princeton) — cite the HAL paper