benchgap
Calibration

HLE w/ tools → AA-AnalystAgent

AA-AnalystAgent is estimated from HLE w/ tools with a offset logistic curve fitted on 5 models measured on both: y = 0.0413 + (0.5651 − 0.0413) / (1 + exp(−26.89·(x − 0.4908))), R² = 1.00, cross-validated error 2.5 pp. It is used for 17 estimates.

Estimated modelHLE w/ toolsAA-AnalystAgentSource
Agents-A147.6%25.2%estimated ± 2.5 pp, medium confidence
Apodex 1.156.1%49.6%estimated ± 2.5 pp, medium confidence
Claude Haiku 5.557.4%51.5%estimated ± 2.5 pp, medium confidence
Claude Sonnet 557.4%51.5%estimated ± 2.5 pp, medium confidence
DeepSeek V4.1 Flash63.9%55.6%estimated ± 2.5 pp, medium confidence
DeepSeek V4 Flash 073145.1%17.5%estimated ± 2.5 pp, medium confidence
DeepSeek V4 Pro 081360.0%53.9%estimated ± 2.5 pp, medium confidence
dots3-note Preview52.6%41.9%estimated ± 2.5 pp, medium confidence
GLM-5.362.5%55.1%estimated ± 2.5 pp, medium confidence
GLM-5.3-Flash55.3%48.2%estimated ± 2.5 pp, medium confidence
Hy4 preview55.4%48.4%estimated ± 2.5 pp, medium confidence
Ornith-1.5-35B-A3B33.4%4.9%estimated ± 2.5 pp, low confidence
Ornith-1.5-397B56.1%49.6%estimated ± 2.5 pp, medium confidence
Ornith-1.5-9B30.5%4.5%estimated ± 2.5 pp, low confidence
Qwen3.7 Max53.5%44.3%estimated ± 2.5 pp, medium confidence
Qwen3.8 Max56.2%49.8%estimated ± 2.5 pp, medium confidence
Step 3.7 Flash47.2%23.8%estimated ± 2.5 pp, medium confidence