Has Step 5 Preview been released?
Yes. StepFun announced it on September 20, 2026 and made it available the same day through its products and its open platform API as model id step-5-preview.
ST
Step 5 Preview
PreviewReasoning modelCoding modelStepFun Step 5 Preview
Also known as: Step5 Preview / Step 5
Release date: 2026-09-20Updated: 2026-09-20Views: 456
Parameters
600B
Context length
1M
Multilingual
Supported
Reasoning ability
4/5
Step 5 Preview is StepFun's flagship foundation model, released on September 20, 2026: a sparse MoE with 600B total and about 27B activated parameters across 92 layers, a 1M-token context window, and text, image and video input. It targets agentic coding and knowledge work, scoring 93.5% on GPQA Diamond, 85.0% on Terminal-Bench v2.1, 88.7% on BrowseComp and 76.0% on MMMU-Pro, with an Artificial Analysis Intelligence Index of 44. It is API-only at launch; StepFun says BF16 weights follow on October 15, 2026.
Data sourced primarily from official releases (GitHub, Hugging Face, papers), then benchmark leaderboards, then third-party evaluators. Learn about our data methodology
Step 5 Preview
Model basics
Reasoning traces
Supported
Thinking modes
Thinking Level · Medium (Default)Thinking Level · LowThinking Level · High
Context length
1M tokens
Max output length
1M tokens
Model type
Reasoning model
Modality (in / out)
Text, Image, Video → Text
Release date
2026-09-20
Model file size
No data
MoE architecture
Yes
Total params / Active params
600B / 27B
Knowledge cutoff
No data
Step 5 Preview
Open source & experience
Code license
Proprietary
Weights license
Proprietary
GitHub repo
N/A
Hugging Face
N/A
Live demo
N/A
Step 5 Preview
Official resources
Paper
DataLearnerAI blog
N/A
Step 5 Preview
API details
API speed
3/5
💡Default unit: $/1M tokens. If vendors use other units, follow their published pricing.
Standard
| Type | Condition | Input | Output |
|---|---|---|---|
| Text | - | $1.00/ 1M tokens | $2.70/ 1M tokens |
Cache PricingPrompt Cache
| Type | TTL | Write | Read |
|---|---|---|---|
| Text | - | — | $0.050/ 1M tokens |
“—” means the modality is not billed in that direction, or the vendor has not published a price for it.
Step 5 Preview
Benchmark Results
Step 5 Preview currently shows benchmark results led by AA-LCR (2 / 171, score 88.30), HLE (9 / 565, score 59.40), GPQA Diamond (25 / 463, score 93.50). This page also consolidates core specs, context limits, and API pricing so you can evaluate the model from benchmark results and deployment constraints together.
General Knowledge
3 evaluationsBenchmark / mode
Score
Rank/total
AI Agent - Tool Usage
6 evaluationsBenchmark / mode
Score
Rank/total
Coding and Software Engineer
6 evaluationsBenchmark / mode
Score
Rank/total
Agent Level Benchmark
4 evaluationsBenchmark / mode
Score
Rank/total
Productivity Knowledge
5 evaluationsBenchmark / mode
Score
Rank/total
Multimodal Understanding
2 evaluationsBenchmark / mode
Score
Rank/total
Compare with other models
- Peer modelStep 5 Preview vs DeepSeek-V4.1-Flash17 benchmarks
- Peer modelStep 5 Preview vs GLM-5.3-Flash15 benchmarks
- Peer modelStep 5 Preview vs Qwen3.8-27B14 benchmarks
- Earlier versionStep 5 Preview vs Step 3.7 Flash8 benchmarks
- Earlier versionStep 5 Preview vs Step 3.5 Flash4 benchmarks
Want a custom combination? Open the compare tool
Step 5 Preview
Publisher
StepFunAI
View publisher details StepFun Step 5 Preview
Model Overview
Released September 20, 2026
Step 5 Preview is StepFun's new flagship foundation model, announced on September 20, 2026 and served the same day through StepFun's products and its open platform API under the model id step-5-preview. StepFun says it is built for real-world agentic work — coding, software engineering, professional knowledge work and financial analysis — where a task needs long context, repeated tool calls and continuous execution rather than a single answer.
Architecture and specifications
StepFun's launch announcement describes a sparse mixture-of-experts model with 600B total parameters and roughly 27B activated per token, built "narrow and deep" across 92 transformer layers. The official platform documentation lists a 1M-token context window, text, image and video input with text output, a maximum output of 1M tokens, and three reasoning-effort levels (low, medium, high). Streaming, tool calling, JSON mode and JSON Schema structured output, and prompt caching are all supported.
Benchmarks
StepFun's published results include GPQA Diamond 93.5%, Terminal-Bench v2.1 85.0%, BrowseComp 88.7% and MMMU-Pro 76.0%. Artificial Analysis independently scores it at 44 on its Intelligence Index, which StepFun cites as a top-three placing among open-weight models.
Weights, access and price
Weights were not published at launch: the model is API-only for now, and StepFun has said it will release BF16 weights on October 15, 2026. The licence that will apply to them has not been stated. StepFun's official price list charges CNY 7 per million input tokens (CNY 0.35 on a cache hit) and CNY 20 per million output tokens; Artificial Analysis recorded the equivalent as roughly $1.00 and $2.70 per million tokens.
Sources
- StepFun: Step 5 Preview official release page — the official announcement
- StepFun Open Platform: Step 5 Preview model documentation — official context window, modalities, limits and features
- StepFun Open Platform: pricing and rate limits — official list price
- Artificial Analysis: Step 5 Preview — third-party Intelligence Index and measured pricing
Step 5 Preview
FAQ
Are Step 5 Preview's weights open?
Not yet. The model is API-only at launch; StepFun has said it will publish BF16 weights on October 15, 2026, and has not stated which licence will apply.
What context window and modalities does Step 5 Preview support?
StepFun's platform documentation lists a 1M-token context window, text, image and video input with text output, and a maximum output of 1M tokens.
How much does Step 5 Preview cost?
StepFun's official price list charges CNY 7 per million input tokens, CNY 0.35 per million cached input tokens, and CNY 20 per million output tokens. Artificial Analysis recorded roughly $1.00 and $2.70 per million input and output tokens.
How capable is Step 5 Preview?
StepFun reports GPQA Diamond 93.5%, Terminal-Bench v2.1 85.0%, BrowseComp 88.7% and MMMU-Pro 76.0%. Artificial Analysis scores it 44 on its Intelligence Index, which StepFun cites as a top-three placing among open-weight models.
DataLearner on WeChat
Follow DataLearner on WeChat for AI model updates and research notes.
