benchgap
Calibration

APEX-Agents-AA → HLE w/ tools

HLE w/ tools is estimated from APEX-Agents-AA with a Hill curve fitted on 5 models measured on both: y = 0.2678 + (0.5705 − 0.2678)·x^6.00 / (0.12372^6.00 + x^6.00), R² = 0.74, cross-validated error 4.1 pp. It is used for 21 estimates.

Estimated modelAPEX-Agents-AAHLE w/ toolsSource
Gemini 3.5 Flash47.1%57.0%estimated ± 4.1 pp, low confidence
Qwen3.8 Max Preview42.4%57.0%estimated ± 4.1 pp, low confidence
GPT-5.6 Terra38.9%57.0%estimated ± 4.1 pp, low confidence
GPT-5.537.7%57.0%estimated ± 4.1 pp, low confidence
GPT-5.6 Luna35.8%57.0%estimated ± 4.1 pp, low confidence
GPT-5.433.3%57.0%estimated ± 4.1 pp, medium confidence
Claude Opus 4.6 (Adaptive)33.0%57.0%estimated ± 4.1 pp, medium confidence
Gemini 3.1 Pro32.0%56.9%estimated ± 4.1 pp, medium confidence
Apodex 1.1 Mini31.2%56.9%estimated ± 4.1 pp, medium confidence
Kimi K2.628.5%56.8%estimated ± 4.1 pp, medium confidence
GPT-5.4 mini28.2%56.8%estimated ± 4.1 pp, medium confidence
Ling 3.0 Flash Fin27.4%56.8%estimated ± 4.1 pp, medium confidence
GPT-5.4 nano24.9%56.6%estimated ± 4.1 pp, medium confidence
Qwen3.7 Plus22.4%56.2%estimated ± 4.1 pp, medium confidence
Grok 4.317.0%53.1%estimated ± 4.1 pp, medium confidence
Qwen3.5 397B15.3%50.4%estimated ± 4.1 pp, medium confidence
Kimi K2.511.5%38.6%estimated ± 4.1 pp, low confidence
MiniMax M2.710.6%35.4%estimated ± 4.1 pp, low confidence
GPT-OSS 120B3.1%26.8%estimated ± 4.1 pp, low confidence
MiMo-V2.5-Pro2.4%26.8%estimated ± 4.1 pp, low confidence
GPT-OSS 20B0.7%26.8%estimated ± 4.1 pp, low confidence

Fitted on

ModelAPEX-Agents-AAHLE w/ tools
Apodex 1.131.2%56.1%
DeepSeek V4 Pro24.3%60.0%
GLM-514.5%50.4%
GLM-5.233.7%54.7%
Step 3.7 Flash14.8%47.2%