benchgap
Calibration

τ³-bench results → AA Agentic Index

AA Agentic Index is estimated from τ³-bench results with a offset logistic curve fitted on 8 models measured on both: y = 0.0493 + (0.1971 − 0.0493) / (1 + exp(−197.23·(x − 0.6683))), R² = 0.66, cross-validated error 5.9 pp. It is used for 9 estimates.

Estimated modelτ³-bench resultsAA Agentic IndexSource
Qwen3.5 397B68.4%19.1%estimated ± 5.9 pp, medium confidence
LFM3-1.7B68.0%18.4%estimated ± 5.9 pp, medium confidence
Pokee-Isaac 28B66.2%8.2%estimated ± 5.9 pp, medium confidence
LFM2.5-24B-A2B8.0%4.9%estimated ± 5.9 pp, low confidence
Mercury 330.2%4.9%estimated ± 5.9 pp, medium confidence
Nemotron 3.5 Bolt 74B-A9B BF1617.0%4.9%estimated ± 5.9 pp, medium confidence
Beam 1.0 501B-A23B FP836.8%4.9%estimated ± 5.9 pp, medium confidence
Beam 1.0 501B-A23B NVFP435.4%4.9%estimated ± 5.9 pp, medium confidence
Spark-X2.6-4B34.7%4.9%estimated ± 5.9 pp, medium confidence

Fitted on

Modelτ³-bench resultsAA Agentic Index
GLM-5.170.6%25.2%
Granite 4.2 8B58.1%3.7%
MiMo-V2.5-Pro72.9%22.7%
MiniMax M2.767.6%16.8%
Mistral Medium 3.5 128B91.4%9.4%
Nemotron 3.5 Lightning 30B A3B NVFP49.5%6.1%
Nemotron 3 Ultra70.9%21.7%
Qwen3.6-35B-A3B67.2%15.0%