StepFun · model
Step 3.7 Flash benchmark scores
As of 2026-10-07, Step 3.7 Flash (StepFun) has measured scores on 23 benchmarks and estimated scores on 92 more.
| Benchmark | Score | Source |
|---|---|---|
| AA-GPQA Diamond | 80.9% | measured |
| AA-HLE | 21.4% | measured |
| AA-SciCode | 43.9% | measured |
| Artificial Analysis Intelligence Index | 19.5% | measured |
| CritPt | 2.3% | measured |
| GDPval-AA | 25.8% | measured |
| GPQA Diamond | 82.2% | estimated ± 2.2 pp, high confidence |
| HLE | 31.7% | estimated ± 9.5 pp, medium confidence |
| IFBench | 66.3% | estimated ± 6.9 pp, medium confidence |
| AA-LCR | 73.7% | measured |
| LiveCodeBench v6 | 87.9% | estimated ± 2.1 pp, high confidence |
| AA-Omniscience Accuracy | 25.8% | measured |
| SciCode | 43.8% | estimated ± 3.2 pp, high confidence |
| Terminal-Bench 2.1 | 59.5% | measured |
| BrowseComp | 75.8% | measured |
| HLE w/ tools | 47.2% | measured |
| IFEval | 93.3% | estimated ± 1.3 pp, high confidence |
| LongBench v2 | 61.5% | estimated ± 5.9 pp, medium confidence |
| VITA-Bench | 43.4% | estimated ± 9.9 pp, low confidence |
| τ²-bench results | 98.5% | measured |
| AA Coding Index | 39.6% | measured |
| AA-MMMU-Pro | 75.3% | measured |
| APEX-Agents | 23.6% | estimated ± 10.0 pp, low confidence |
| APEX-Agents-AA | 14.8% | measured |
| BioMysteryBench (human-difficult) | 31.9% | estimated ± 5.9 pp, low confidence |
| DeepSearchQA | 92.8% | measured |
| SWE-bench Verified | 79.6% | estimated ± 2.5 pp, high confidence |
| AutomationBench | 28.4% | estimated ± 9.7 pp, low confidence |
| BFCL v4 | 74.3% | estimated ± 2.6 pp, high confidence |
| CyberGym | 76.8% | estimated ± 2.4 pp, medium confidence |
| JobBench | 36.0% | estimated ± 7.3 pp, low confidence |
| SWE-bench Pro | 56.3% | measured |
| τ³-bench results | 71.0% | estimated ± 2.1 pp, low confidence |
| WideResearch | 76.5% | estimated ± 3.4 pp, medium confidence |
| AA Agentic Index | 19.8% | estimated ± 2.8 pp, high confidence |
| MMLU-Pro | 85.6% | estimated ± 3.5 pp, high confidence |
| GPQA | 84.8% | estimated ± 3.2 pp, high confidence |
| AA-IFBench | 67.3% | measured |
| Gert Labs | 51.6% | measured |
| AA-AnalystAgent | 23.8% | estimated ± 2.5 pp, medium confidence |
| AA EnterpriseOps-Gym | 52.8% | estimated ± 3.6 pp, medium confidence |
| AA Harvey LAB | 86.6% | estimated ± 1.7 pp, medium confidence |
| ApprenticeBench | 0.0% | estimated ± 3.1 pp, low confidence |
| ARC-AGI-1 | 46.2% | estimated ± 5.4 pp, low confidence |
| ARC-AGI-2 | 15.2% | estimated ± 14.2 pp, low confidence |
| CursorBench 3.1 | 29.0% | estimated ± 3.4 pp, low confidence |
| CursorBench 3.2 | 52.6% | estimated ± 3.6 pp, high confidence |
| Terminal-Bench 3.0 | 5.2% | estimated ± 9.1 pp, low confidence |
| FrontierCode 1.1 Main | 14.7% | estimated ± 3.8 pp, medium confidence |
| FrontierSWE v2 | 8.6% | estimated ± 9.0 pp, medium confidence |
| OfficeQA Pro | 50.3% | estimated ± 7.2 pp, low confidence |
| OSWorld-Verified | 67.1% | estimated ± 5.4 pp, medium confidence |
| Vals GPQA Diamond | 73.6% | estimated ± 7.4 pp, medium confidence |
| Vals LiveCodeBench | 82.9% | estimated ± 4.7 pp, medium confidence |
| Vals MMLU-Pro | 84.2% | estimated ± 2.2 pp, high confidence |
| Vals SWE-bench | 62.6% | estimated ± 5.4 pp, medium confidence |
| Terminal-Bench 2.1 (Vals AI) | 51.9% | estimated ± 7.4 pp, medium confidence |
| AA AutomationBench | 4.6% | estimated ± 5.7 pp, low confidence |
| GDP.pdf | 9.4% | estimated ± 5.4 pp, medium confidence |
| AA ITBench | 58.6% | estimated ± 3.4 pp, low confidence |
| AA Tau3 Banking | 22.0% | estimated ± 5.8 pp, medium confidence |
| AA Terminal-Bench 2.1 | 53.4% | estimated ± 3.5 pp, high confidence |
| AA Terminal-Bench 4.0 | 0.0% | estimated ± 8.9 pp, medium confidence |
| CursorBench 4.0 | 10.9% | estimated ± 1.6 pp, medium confidence |
| CWE-bench v1 | 35.7% | estimated ± 6.9 pp, low confidence |
| DeepSWE | 3.9% | estimated ± 6.4 pp, low confidence |
| HLE w/o tools | 18.4% | estimated ± 5.7 pp, medium confidence |
| OSWorld 2.0 | 0.9% | estimated ± 4.5 pp, low confidence |
| PostTrainBench v1.1 | 23.4% | estimated ± 4.7 pp, medium confidence |
| Toolathlon-Verified | 50.4% | estimated ± 2.4 pp, medium confidence |
| Vibe Code Bench | 14.6% | estimated ± 10.3 pp, low confidence |
| CharXiv | 80.3% | estimated ± 4.9 pp, high confidence |
| CharXiv w/o tools | 79.8% | estimated ± 2.4 pp, high confidence |
| ExploitGym | 0.0% | estimated ± 6.1 pp, low confidence |
| C-Eval | 91.0% | estimated ± 1.0 pp, medium confidence |
| Claw-Eval | 67.1% | measured |
| MathVision | 87.2% | estimated ± 2.7 pp, high confidence |
| MCP Atlas | 73.5% | estimated ± 7.3 pp, medium confidence |
| MCP-Tasks | 86.9% | estimated ± 6.4 pp, low confidence |
| MMLU-Redux | 94.7% | estimated ± 1.1 pp, medium confidence |
| MMMU-Pro | 81.8% | estimated ± 2.4 pp, high confidence |
| NL2Repo | 28.0% | estimated ± 5.1 pp, low confidence |
| QwenClawBench | 55.9% | estimated ± 6.0 pp, low confidence |
| ScreenSpot Pro | 81.9% | estimated ± 2.5 pp, medium confidence |
| SuperGPQA | 70.8% | estimated ± 7.5 pp, medium confidence |
| Toolathlon | 49.5% | measured |
| V* | 95.3% | measured |
| VideoMMMU | 87.0% | estimated ± 2.2 pp, low confidence |
| ERQA | 63.8% | estimated ± 3.9 pp, high confidence |
| HealthBench Hard | 6.6% | estimated ± 7.7 pp, low confidence |
| LiveCodeBench Pro | 80.8% | estimated ± 6.0 pp, low confidence |
| MedXpertQA (MM) | 72.6% | estimated ± 6.5 pp, low confidence |
| MMLU-Pro (Arcee) | 85.8% | estimated ± 4.0 pp, medium confidence |
| React Native Evals | 74.0% | estimated ± 3.5 pp, high confidence |
| ResearchClawBench | 18.2% | estimated ± 1.3 pp, high confidence |
| SWE-Rebench | 0.1% | estimated ± 8.9 pp, low confidence |
| ARC-AGI-3 | 1.7% | estimated ± 3.4 pp, medium confidence |
| BioMysteryBench (human-solvable) | 81.7% | estimated ± 0.6 pp, low confidence |
| DRACO | 72.1% | estimated ± 3.0 pp, medium confidence |
| FrontierCode 1.1 Extended | 54.7% | estimated ± 3.0 pp, low confidence |
| HealthBench (raw) | 29.6% | estimated ± 6.4 pp, low confidence |
| HealthBench (length-adjusted) | 53.9% | estimated ± 3.7 pp, low confidence |
| HealthBench Professional | 57.7% | estimated ± 4.8 pp, low confidence |
| HealthBench Professional (raw) | 41.5% | estimated ± 5.0 pp, low confidence |
| HLE-Verified | 15.3% | estimated ± 2.0 pp, low confidence |
| LABBench2 | 64.0% | estimated ± 1.8 pp, low confidence |
| MMMU | 83.2% | estimated ± 0.9 pp, high confidence |
| LiveCodeBench | 72.9% | estimated ± 8.9 pp, low confidence |
| OpenHarmony Bench | 51.0% | estimated ± 2.4 pp, medium confidence |
| SimpleVQA | 79.2% | measured |
| ZeroBench | 17.8% | estimated ± 6.4 pp, low confidence |
| MMLU | 89.3% | estimated ± 1.1 pp, medium confidence |
| MMMU-Pro w/ Python | 78.8% | estimated ± 1.6 pp, high confidence |
| RealWorldQA | 85.1% | estimated ± 1.2 pp, medium confidence |
| OmniDocBench 1.5 | 89.6% | estimated ± 6.4 pp, low confidence |