- GPQA Diamond 93.5
- Best of HLE 46.5 / HLE w/ tools (text-only) 59.4 59.4
- AA-LCR v1.1 88.3
- CritPt 20.9
- DeepSWE v1.1 (SWE-agent) 67.7
- Terminal-Bench v2.1 85
- Terminal-Bench v4 33.3
- CyberGym 84.7
- SciCode 58.9
- RoadmapBench 54.3
- ProgramBench (Pass Rate) 80.5
- SWE-Marathon v1.1 (Partial Score) 72.7
- MLS-Bench-Lite 40.5
- Mean of SWE-Atlas-QnA 63.6 / SWE-Atlas-Test-writing 50.8 57.2
- StepCodeBench (avg@4, internal) 49
- GDPval-AA v2.1 (Artificial Analysis, 2026-09-20) 1566
- τ³-Banking 42.5
- AutomationBench (public); AutomationBench-AA 51.0 44
- AA-Briefcase v1.1 1433
- Toolathlon-Verified 74.1
- MCP-Atlas 85.6
- PresentBench 76.8
- OfficeQA Pro 60.3
- SpeadSheet v2 29.4
- JobBench 59
- Apex-Agents 37.8
- Draco 83.3
- BrowseComp 88.7
- Mean of FinStepBench-LiveSearch 74.5 / FinStepBench-CorporateValuation 60.6 / FinStepBench-FinanceDR 55.8 63.6
- FrontierFinance 66.4
- Agents Last Exam (ALE-CLI) 29.5
- MMMU-Pro 76
- GDP.pdf 14.8
Step 5 Preview
Step 5 Preview ranks first on 2 of the 71 benchmarks with five or more models scored.
76
Benchmarks with a score
2
Ranked first, out of five or more models
6
In the top three
StepFun Sep 2026 Step 5 Preview
Closed weights
160.1
#30 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- MMMU-Pro 76% #36 /90
- SciCode 58.9% #3 /49
- APEX-Agents 37.8% #4 /12
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 | |
|---|---|---|---|---|
| Humanity's Last Exam | 59.4% | #15 of 154 | Lab model card | |
| AA-LCR 1.1 | 88.3% | #1 of 6 | Lab model card | |
| LMArena (Elo) | 1,448 | #65 of 335 | Official leaderboard | |
| Extended NYT Connections | 76.6% | #50 of 103 | Official leaderboard | |
| StudyArena Text | 1,131 | #84 of 85 | Official leaderboard |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| GPQA Diamond | 93.5% | #11 of 302 | Lab model card | |
| CritPt | 20.9% | #4 of 13 | Lab model card | |
| BioMysteryBench | 67.4% | #20 of 30 | Official leaderboard | |
| ProofBench v1.1 | 42% | #35 of 52 | Official leaderboard | |
| MysteryMechanism | 18.5% | #28 of 31 | Official leaderboard |
Coding
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SciCode | 58.9% | #3 of 49 | Lab model card | |
| SWE-Marathon | 72.7% | #1 of 11 | Lab model card | |
| WebDev Arena | 1,570 | #26 of 105 | arena.ai, via Epoch AI | |
| Code Arena | 1,573 | #41 of 128 | Official leaderboard | |
| Vibe Code Bench | 70.4% | #49 of 120 | Official leaderboard | |
| Code Migration | 39% | #32 of 78 | Official leaderboard | |
| Doc2Harness | 40.8% | #8 of 19 | Official leaderboard | |
| DeepSWE | 67.7% | #24 of 56 | Lab model card | |
| Image-to-WebDev Arena | 1,551 | #26 of 55 | Official leaderboard | |
| SWE Atlas | 57.2% | #8 of 14 | Lab model card | |
| Tencent WorkBuddy Bench | 75.6% | #8 of 14 | Official leaderboard | |
| Code Migration v1.1 | 47% | #18 of 30 | Official leaderboard | |
| ProgramBench tests passed | 80.5% | #6 of 10 | Lab model card | |
| StepCodeBench | 49% | #3 of 5 | Official leaderboard | |
| RSI-Exam | 35.2% | #13 of 18 | Official leaderboard | |
| MLS-Bench | 40.5% | #4 of 5 | Lab model card | |
| IOI v2 (Vals) | 38.1% | #38 of 45 | Official leaderboard | |
| FrontierCS 2 | 10.4% | #17 of 19 | Official leaderboard | |
| RSI Arena | 934 | #9 of 10 | Official leaderboard | |
| RoadmapBench | 54.3% | #1 of 1 | Lab model card |
Agents and tools
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MCP Atlas | 85.6% | #5 of 46 | Lab model card | |
| τ³-bench Banking | 42.5% | #3 of 22 | Lab model card | |
| BrowseComp | 88.7% | #13 of 81 | Lab model card | |
| Terminal-Bench 2.1 | 85% | #17 of 56 | Lab model card | |
| AutomationBench | 44% | #20 of 47 | Lab model card | |
| Toolathlon-Verified | 74.1% | #12 of 21 | Lab model card | |
| DRACO | 83.3% | #7 of 12 | Lab model card | |
| Agent Arena | -3.74% | #39 of 56 | Official leaderboard | |
| Hermes Bench | 69.2% | #13 of 17 | Official leaderboard | |
| Terminal-Bench 4.0 | 33.3% | #22 of 28 | Lab model card |
Professional work
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Harvey LAB (Legal Agent Benchmark) | 9.17% | #22 of 80 | Official leaderboard | |
| APEX-Agents | 37.8% | #4 of 12 | Lab model card | |
| FinStepBench DeepResearch | 55.8% | #2 of 6 | Official leaderboard | |
| Legal Research Bench | 40.4% | #37 of 93 | Official leaderboard | |
| Harvey LAB v1.1 (HLAB-Vals-1.1) | 12.9% | #17 of 38 | Official leaderboard | |
| Tax Agent Bench | 27.3% | #40 of 84 | Official leaderboard | |
| Excel Modeling Benchmark v1.1 | 69.5% | #20 of 42 | Official leaderboard | |
| Vals Legal Index | 70.2% | #19 of 38 | Official leaderboard | |
| FrontierFinance | 66.4% | #1 of 2 | Lab model card | |
| Vals Finance Agent v2 | 50.7% | #42 of 80 | Official leaderboard | |
| Agents' Last Exam | 29.5% | #14 of 26 | Lab model card | |
| Excel Modeling Benchmark | 57.1% | #41 of 76 | Official leaderboard | |
| Vals Index v2.1 | 49.4% | #32 of 49 | Official leaderboard | |
| JobBench | 59% | #10 of 15 | Lab model card | |
| FinStepBench CorporateValuation | 60.6% | #4 of 6 | Official leaderboard | |
| FinStepBench LiveSearch | 74.5% | #4 of 6 | Official leaderboard | |
| PresentBench | 76.8% | #2 of 3 | Lab model card | |
| Vals Index v2.2 | 54.8% | #25 of 37 | Official leaderboard | |
| Vals Finance Agent v2.1 | 58.9% | #29 of 40 | Official leaderboard | |
| Public Benefits Bench | 58.1% | #37 of 50 | Official leaderboard | |
| Vals Finance Index | 55.4% | #30 of 40 | Official leaderboard | |
| OfficeQA Pro | 60.3% | #12 of 16 | Lab model card | |
| Investor's Last Exam | 55.1 | #8 of 10 | Official leaderboard | |
| GDPval-AA v2.1 | 1,566 | #9 of 10 | Lab model card | |
| AA-Briefcase v1.1 | 1,433 | #8 of 8 | Lab model card | |
| SpreadsheetBench 2 | 29.4% | #3 of 3 | Lab model card | |
| FinStepBench | 63.6% | #1 of 1 | Lab model card |
Vision, audio and robotics
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| Vision Arena | 1,245 | #50 of 136 | Official leaderboard | |
| MMMU-Pro | 76% | #36 of 90 | Lab model card | |
| GDP.pdf | 14.8% | #16 of 17 | Lab model card |
Safety and security
Other
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| SuperCLUE | 41.2 | #11 of 16 | Official leaderboard | |
| LiveBench | 54.9% | #25 of 31 | Official leaderboard | |
| Vals Science Index | 29.6% | #21 of 25 | Official leaderboard | |
| Hermes Index | 32.8 | #15 of 17 | Official leaderboard |
Its lab's numbers put it at 160.1, number 30 of 404
The Self-Reported Index fits one score to the 22 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 157.5 to 162.7. How we compute it.
- GPQA Diamond 93.5
- BrowseComp 88.7
- MCP Atlas 85.6
- Terminal-Bench 2.1 85
- DRACO 83.3
- ProgramBench tests passed 80.5
- MMMU-Pro 76
- Toolathlon-Verified 74.1
- SWE-Marathon 72.7
- DeepSWE 67.7
- OfficeQA Pro 60.3
- Humanity's Last Exam 59.4
- JobBench 59
- SciCode 58.9
- SWE Atlas 57.2
- AutomationBench 44
- τ³-bench Banking 42.5
- APEX-Agents 37.8
- Terminal-Bench 4.0 33.3
- Agents' Last Exam 29.5
- CritPt 20.9
- GDP.pdf 14.8
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 |
|---|---|---|---|
| StepCodeBench | 49% StepCodeBench (avg@4, internal) | 49% Leaderboard, avg@4; high | Same |
Its launch post
The benchmarks the lab chose to headline, with the scores it gave.