StepFun · Step 5
Step 5 Preview
StepFun's flagship 600B/27B-active MoE agentic reasoning model with 1M context and text/image/video input, API-only in preview with open weights promised for mid-October 2026.
- AA Intelligence
- 43.7#15 of 142
- Input / output per 1M
- $1.00/ $2.70
- Output speed
- 77tok/s
- Context window
- 1M tokens
Composite indexes
- Artificial Analysis Intelligence Index#15 of 14244
Leader: Claude Opus 5.5 at 58
Coding
- Terminal-Bench 4.0#14 of 8833.3%
Leader: Claude Mythos 5.1 at 60.9%
- SciCode#8 of 9458.9%
Leader: Claude Opus 5.5 at 66.9%
Agents and tool use
- GDPval-AA#16 of 1031566
Leader: Claude Opus 5.5 at 1846
Reasoning and knowledge
- Humanity’s Last Exam#17 of 14746.5%
Leader: Claude Opus 5.5 at 61.4%
- CritPt#14 of 13920.9%
Leader: GPT-5.6 Sol at 32.3%
Vision and long context
- AA Long Context Reasoning#2 of 13988.3%
Leader: Kimi K3 at 88.7%
- MMMU-Pro#31 of 7276.4%
Leader: Claude Opus 5.5 at 87.7%
Other published results
| Result | Score | Reported by | Source |
|---|---|---|---|
| AA-Omniscience Index · AA | 16.4 index (-100..100) | independent | Artificial Analysis: Step 5 Preview ↗ |
| DeepSWE v1.1 | 67.7% | vendor | MarkTechPost: StepFun launches Step 5 Preview (StepFun-reported figures) ↗ |
| FrontierFinance · high effort | 66.4% | vendor | MarkTechPost: StepFun launches Step 5 Preview (StepFun-reported figures) ↗ |
| DRACO · high effort | 83.3% | vendor | MarkTechPost: StepFun launches Step 5 Preview (StepFun-reported figures) ↗ |
| StepCodeBench | 49% | vendor | MarkTechPost: StepFun launches Step 5 Preview (StepFun-reported figures) ↗ |
| ProgramBench · as reported by StepFun; metric may differ from other labs' ProgramBench figures | 80.5% | vendor | MarkTechPost: StepFun launches Step 5 Preview (StepFun-reported figures) ↗ |