- HLE w. tool (acc) 47.2% (text-only 49.7%) 47.2
- BrowseComp (acc) 75.8
- deepsearchQA (F1), acc 81.7 92.8
- ResearchRubrics (score) 71.7
- Toolathlon (internal fixed version) 49.5
- Claweval-v1.1 (pass^3) 67.1
- GDPval-Stirrup 1415.8 (ii 45.8%) 1415.8
- SWE-MTLG 72.4
- SWE-Bench Pro 56.3
- SWE-Bench Verified 76.5
- Terminal-Bench 2.1 59.6
- AA-LCR (avg@16/acc) 63.9
- SimpleVQA (with search) 79.2
- WorldVQA 58.1
- BC-VL 59
- V* (with Python) 95.3
- Mean of HR-Bench 4K 89.13 / HR-Bench 8K 86.34 (with Python) 87.7
- VisualProbe (with Python) 65
- Score of Android Daily Benchmark 61.9
- Step-SWE-Bench avg over six harnesses 67.1
Step 3.7 Flash
Step 3.7 Flash places in the top three on 5 benchmarks.
42
Benchmarks with a score
0
Ranked first, out of five or more models
5
In the top three
StepFun May 2026 Step 3.7 Flash
Open weights
152.5
#78 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- AA-LCR (Artificial Analysis Long Context Reasoning) 63.9% #19 /38
- SimpleVQA 79.2% #2 /30
- Claw-Eval 80.5% #3 /24
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SimpleVQA | 79.2% | #2 of 30 | Lab model card | |
| Humanity's Last Exam | 47.2% | #53 of 154 | Lab model card | |
| AA-LCR (Artificial Analysis Long Context Reasoning) | 63.9% | #19 of 38 | Lab model card | |
| DesignArena ASCII Art | 1,142 | #51 of 77 | Official leaderboard | |
| Extended NYT Connections | 39.7% | #76 of 103 | Official leaderboard | |
| Judgemark v4 | 28.1 | #36 of 40 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MathArena Apex Shortlist | 76.6% | #12 of 38 | Official leaderboard | |
| MathArena (Apex, IMO, USAMO) | 14.6% | #17 of 47 | Official leaderboard | |
| HMMT 2026 | 87.9% | #18 of 31 | Official leaderboard | |
| AIME 2026 | 95% | #19 of 31 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SWE-bench Verified | 76.5% | #38 of 155 | Lab model card | |
| SWE-bench Multilingual | 72.4% | #39 of 73 | Lab model card | |
| DesignArena Website | 1,187 | #82 of 151 | Official leaderboard | |
| SWE-Bench Pro | 56.3% | #44 of 81 | Lab model card | |
| DesignArena Code | 1,170 | #83 of 143 | Official leaderboard | |
| DesignArena UI Component | 1,168 | #81 of 139 | Official leaderboard | |
| ALE-Bench | 1,277 | #76 of 126 | Official leaderboard | |
| DesignArena Game Dev | 1,144 | #87 of 143 | Official leaderboard | |
| DesignArena Data Visualization | 1,173 | #77 of 118 | Official leaderboard | |
| DesignArena 3D Design | 1,111 | #89 of 136 | Official leaderboard | |
| DesignArena SVG | 1,074 | #77 of 103 | Official leaderboard | |
| DesignArena Image to HTML | 1,112 | #48 of 51 | Official leaderboard | |
| Monthly-SWEBench | 22% | #12 of 12 | Official leaderboard | |
| Step-SWE-Bench | 67.1% | #1 of 1 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ClawProBench | 68.4% | #6 of 49 | Official leaderboard | |
| Claw-Eval | 80.5% | #3 of 24 | Official leaderboard | |
| DeepSearchQA | 92.8% | #4 of 23 | Lab model card | |
| Claw-Eval Pass^3 | 67.1% | #3 of 13 | Lab model card | |
| BrowseComp-VL | 59% | #1 of 3 | Lab model card | |
| Toolathlon | 49.5% | #13 of 29 | Lab model card | |
| ResearchRubrics | 71.7% | #3 of 6 | Lab model card | |
| AndroidDaily | 61.9% | #1 of 2 | Lab model card | |
| BrowseComp | 75.8% | #41 of 81 | Lab model card | |
| SnakeBench | 21 | #119 of 188 | Official leaderboard | |
| GameHorizon | 64.1 | #25 of 37 | Official leaderboard | |
| PACT | 1,500 | #19 of 26 | Official leaderboard | |
| Terminal-Bench 2.1 | 59.6% | #46 of 56 | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GDPval-AA | 1,416 | #12 of 16 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| WorldVQA | 58.1% | #2 of 9 | Lab model card | |
| V* | 95.3% | #4 of 17 | Lab model card | |
| HR-Bench | 87.7% | #2 of 2 | Lab model card | |
| VisualProbe | 65% | #1 of 1 | Lab model card |
Its lab's numbers put it at 152.5, number 78 of 404
The Self-Reported Index fits one score to the 13 capability results StepFun reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 150.0 to 155.0. How we compute it.
- V* 95.3
- DeepSearchQA 92.8
- SimpleVQA 79.2
- SWE-bench Verified 76.5
- BrowseComp 75.8
- SWE-bench Multilingual 72.4
- Claw-Eval Pass^3 67.1
- AA-LCR (Artificial Analysis Long Context Reasoning) 63.9
- Terminal-Bench 2.1 59.6
- WorldVQA 58.1
- SWE-Bench Pro 56.3
- Toolathlon 49.5
- Humanity's Last Exam 47.2
Its lab's number next to a leaderboard
Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.
| Benchmark | Lab says | Board says | Gap |
|---|---|---|---|
| Claw-Eval Pass^3 | 67.1% Claweval-v1.1 (pass^3) | 68.3% Leaderboard | −1.2 pts |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.