benchgap
Calibration

HLE w/ tools → AA ITBench

AA ITBench is estimated from HLE w/ tools with a linear curve fitted on 5 models measured on both: y = -0.8762·x + 0.9993, R² = 0.82, cross-validated error 3.4 pp. It is used for 17 estimates.

Estimated modelHLE w/ toolsAA ITBenchSource
Agents-A147.6%58.2%estimated ± 3.4 pp, low confidence
Apodex 1.156.1%50.8%estimated ± 3.4 pp, medium confidence
Claude Haiku 5.557.4%49.6%estimated ± 3.4 pp, medium confidence
Claude Opus 564.7%43.2%estimated ± 3.4 pp, medium confidence
Claude Sonnet 557.4%49.6%estimated ± 3.4 pp, medium confidence
Claude Sonnet 5.564.5%43.4%estimated ± 3.4 pp, medium confidence
DeepSeek V4 Flash 073145.1%60.4%estimated ± 3.4 pp, low confidence
DeepSeek V4 Pro 081360.0%47.4%estimated ± 3.4 pp, medium confidence
dots3-note Preview52.6%53.8%estimated ± 3.4 pp, low confidence
Hy4 preview55.4%51.4%estimated ± 3.4 pp, medium confidence
Nemotron 3 Ultra37.4%67.2%estimated ± 3.4 pp, low confidence
Ornith-1.5-35B-A3B33.4%70.7%estimated ± 3.4 pp, low confidence
Ornith-1.5-397B56.1%50.8%estimated ± 3.4 pp, medium confidence
Ornith-1.5-9B30.5%73.2%estimated ± 3.4 pp, low confidence
Qwen3.7 Max53.5%53.1%estimated ± 3.4 pp, low confidence
Qwen3.8 Max56.2%50.7%estimated ± 3.4 pp, medium confidence
Step 3.7 Flash47.2%58.6%estimated ± 3.4 pp, low confidence