Benchmark
Humanity's Last Exam (with tools)
Scores are percentages. Higher is better.
- Models and systems measured
- 61
- Scores recorded
- 113
- Sources
- 2
- IQ weight
- 1.73
Score distribution
- 51.4%
- Mean
- 54%
- Median
- 13.5%
- Std dev
- 36.5%
- 10th percentile
- 63.9%
- 90th percentile
- 6.3%
- Min
This benchmark has a weight of 1.73 for chat models, measured over 14 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.
Leaderboard
| # | Model or system | Score |
|---|---|---|
| 101 | Claude Sonnet 5search/fetch/code · low effort · 1M total tokens· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 36.5% |
| 102 | Claude Sonnet 4.5web search/fetch + code · no reasoning · decontaminated · Sonnet 4.5 grader· Reported by Anthropic· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 28.4% |
| 103 | Gemma 4 31B ITGemma 4 Technical Report · Table 5 · thinking mode unless noted · with search· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 26.5% |
| 104 | Nemotron 3 Super 120B A12BNeMo Evaluator / mostly NeMo Skills · with tools· Reported by NVIDIA· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 22.8% |
| 105 | Nemotron 3 Super 120B A12Bofficial blog Figure 3 · rounded · +tools· Reported by NVIDIA Developer· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 22.8% |
| 106 | GPT OSS 120Bofficial model card Table 3 · reasoning high · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 19% |
| 107 | GPT OSS 20bofficial model card Table 3 · reasoning high · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 17.3% |
| 108 | Gemma 4 26B A4BGemma 4 Technical Report · Table 5 · thinking mode unless noted · with search· Reported by Google DeepMind· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 17.2% |
| 109 | GPT OSS 120Bofficial model card Table 3 · reasoning medium · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 11.3% |
| 110 | GPT OSS 120Bofficial model card Table 3 · reasoning low · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 9.1% |
| 111 | GPT OSS 20bofficial model card Table 3 · reasoning medium · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 8.8% |
| 112 | GPT OSS 20bofficial model card Table 3 · reasoning low · with tools; Table 3 does not name the exact tool · pass@1· Reported by OpenAI· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 6.3% |
Showing 101–112 of 112 results.
Not ranked
The source marks these results as descriptive, not capability scores, so we list them by name instead of ranking them.
| Model or system | Score |
|---|---|
| Muse Spark (version unspecified)Meta Model API · highest effort · version unspecified · full 2,500 questions · bash + browser tools · o3-mini judge· Reported by Meta, Meta Superintelligence Labs· via BenchBoard docs· Reported by the model's developer. The source relaying it did not verify it. | 50.4% |
Data credits: Data by LLM Stats · 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.