StepFun AI Review 2026: Step Plan, API Pricing, Open Weights Guide
Chinese AI company behind Step 5 Preview (open weights), Step 3.7 Flash for agents and the StepAudio 3 voice models, with an API and coding plan.
StepFun: Step 5 Preview, Step Flash Models and Voice AI
StepFun is a Chinese AI company founded in April 2023 that builds the Step models. Its frontier model, Step 5 Preview, targets complex work and agent tasks, and its weights are available on Hugging Face and GitHub. Step 3.7 Flash and Step 3.5 Flash are faster models for agents, and the StepAudio 3 family covers real-time voice, speech synthesis, voice and music generation, and speech recognition. You can try models in AI Studio or use them through the StepFun API.
StepFun sits alongside other Chinese AI providers covered on this site, including DeepSeek, Qwen, Zhipu AI and Kimi AI. See our roundups of Chinese AI chatbots and free Chinese AI tools, and read is DeepSeek safe to use for data questions. For comparisons, see Claude and our AI chatbots and assistants category.
Quick Navigation
Quick Summary
Overall Rating 4.2/5 Best For Developers building agents, coding tools and voice apps Ease of Use 4.0/5 Business Value 4.3/5
What Is StepFun and Why Does It Matter?
StepFun is the company behind the Step series of models. Its line-up runs from Step 5 Preview, a frontier model for complex and agentic work with open weights, to Step 3.7 Flash, a multimodal model built for agent workflows, and Step 3.5 Flash, which focuses on speed. The StepAudio 3 family adds real-time voice interaction, text-to-speech, voice generation, music generation and speech recognition. Developers can try models in AI Studio, call them through the API at api.stepfun.ai, or subscribe to a monthly coding plan for Step Code.
Open source: Yes. StepFun has released Step 5 Preview model weights on Hugging Face and GitHub. (Checked on Hugging Face and official sites, November 2026.)
Who Should Use StepFun?
- Developers building agentic workflows Step 3.7 Flash is optimized for real Agent tasks with stable tool calling and deep compatibility with mainstream Agent frameworks.
- Teams needing multimodal document parsing Native vision understanding handles PDFs, tables, charts, and UI screenshots in one pass, eliminating the need for vision MCP middleware.
- Enterprises requiring deployable open-weight models Step 3.7 Flash offers open weights for customization, fine-tuning, and on-premise deployment.
- Cost-sensitive production environments The Flash model reduces latency and call overhead, making it suitable for high-volume business workflows.
Professional reality: StepFun is a Chinese company. If you use its hosted API or apps, read its privacy terms before sending personal or business data. For sensitive work, self-hosting the open Step 5 Preview weights keeps data on your own hardware.
StepFun Features That Drive Business Results
Step 3.7 Flash
A multimodal Flash model built for real Agent workflows — it sees, thinks clearly, and acts. Handles office tasks, coding, and Agent pipelines with stable tool calling.
Deep compatibility with mainstream Agent frameworks for reliable execution.
Flagship Capabilities, Fully Open
Step 3.7 Flash offers flagship-level research, document parsing, and coding abilities with open weights — deployable, customizable, and fine-tunable.
Full control over deployment and customization for enterprise needs.
Native Multimodal Understanding
Parses PDFs, tables, charts, and UI screenshots in one pass — no need for vision MCP intermediaries.
Saves API calls and reduces latency in multimodal pipelines.
Step 3.5 Flash
A language model designed for Agents — extremely fast, with rapid response times for high-throughput scenarios.
Accelerates agentic applications where speed is critical.
Step 2 & Step 1
Step 2 is a trillion-parameter self-developed base with deep reasoning and precise multi-level instruction following. Step 1 enables single-card low-bit long-context inference.
Scales from massive reasoning tasks to efficient deployment on limited hardware.
Step API & Studio
The open platform provides stable, high-performance, easy-to-integrate APIs. Studio lets you define your own AI work partner across Mac, Windows, iOS, Android, and Web.
Build and deploy AI applications with minimal friction.
StepFun Pricing in 2026
StepFun sells API access to its models on its API platform and offers a monthly subscription, Flash Pro, for coding and agent work with Step Code. Specific prices are shown on the API platform rather than the homepage. Step 5 Preview weights are free to download.
| Plan | Price | What You Get |
|---|---|---|
| Flash Mini | $6.99 per month | 400M Credits monthly; $18.99 quarterly, $69.99 yearly. |
| Flash Plus | $9.99 per month | 1,600M Credits monthly; $26.99 quarterly, $95.99 yearly. |
| Flash Pro (monthly) Best Value | $29 per month | One plan for coding, development and agent workflows |
| Flash Max | $99 per month | 40,000M Credits monthly; $269 quarterly, $989 yearly. |
If you are comparing Chinese AI providers by cost, you can read our DeepSeek review and our Qwen review for a closer look at how their offerings stack up.
Where StepFun Is Strong / Where It Needs Care
- Open frontier modelStep 5 Preview weights are on Hugging Face and GitHub, so you can run it yourself.
- Fast Flash modelsStep 3.7 Flash and 3.5 Flash are tuned for agents and quick responses.
- Strong voice modelsStepAudio 3 covers real-time voice, speech synthesis, voice and music generation, and speech recognition.
- Several ways inAI Studio, the API platform and a monthly coding plan suit different users.
- Prices not on the homepageYou need to visit the API platform to see current prices.
- Mainly for developersMost value comes through the API and coding plan rather than a consumer app.
- Preview statusStep 5 is still labelled a preview.
- PrivacyHosted use sends data to a Chinese company; self-host for sensitive work.
Real-World Use Cases
Agent Workflows
Use Step 3.7 Flash for agents that call tools to finish office and coding tasks.
Document Parsing
Parse PDFs, tables, charts and UI screenshots in one pass with Step 3.7 Flash.
Coding With Step Code
Use the Flash Pro plan and Step Code for development and agent work.
Voice Apps
Build voice assistants and audio features with StepAudio 3 Realtime, TTS and ASR.
How to Get Started With StepFun
Visit stepfun.com and open AI Studio to try the models.
Create an account on the API platform and generate an API key.
Call https://api.stepfun.ai/v1/chat/completions with a model such as step-3.7-flash.
For self-hosting, download Step 5 Preview from Hugging Face or GitHub.
Is StepFun Worth It in 2026?
StepFun is worth a look for developers who want capable Chinese models for agents, coding and voice. Step 5 Preview with open weights is the standout, the Flash models are quick, and StepAudio 3 is a strong voice family. It is less suited to everyday consumers, prices sit on the API platform rather than the homepage, and hosted use sends data to a Chinese company.
StepFun vs the Competition
| Decision Area | StepFun | When Another Option Wins |
|---|---|---|
| Models | Step 5 Preview, Step 3.7 Flash and 3.5 Flash, plus the StepAudio 3 voice family. | If you need a 1M-token context, DeepSeek or Kimi AI offer that. |
| Open weights | Step 5 Preview weights are on Hugging Face and GitHub. | If you want a larger open-model ecosystem, Qwen may be more established. |
| API | OpenAI-style chat completions API at api.stepfun.ai. | If you already rely on another provider's API, switching takes some work. |
| Products | AI Studio, the API platform and the Step Code coding plan. | If you want a simple consumer chat app, Doubao or DeepSeek may be easier. |
| Voice | StepAudio 3 for real-time voice, TTS, voice and music generation, and ASR. | If you only need text, a text-only API may be cheaper. |
StepFun vs DeepSeek
DeepSeek is known for strong reasoning, a free chat and very low API prices.
Choose StepFun if: You want open frontier weights, agent-ready Flash models and voice models from one provider. Choose DeepSeek if: You mainly need cheap text reasoning with a 1M-token context.
StepFun vs Qwen
Qwen is Alibaba's model family, with many sizes and a large open-source community.
Choose StepFun if: You want Step 5 Preview and the StepAudio voice models. Choose Qwen if: You want Qwen's wider range of model sizes and its consumer chat app.
Frequently Asked Questions
What is StepFun?
StepFun is a Chinese AI company founded in April 2023. It builds the Step models, including Step 5 Preview, Step 3.7 Flash and the StepAudio 3 voice family.
Which models does StepFun offer?
Step 5 Preview (frontier, open weights), Step 3.7 Flash and Step 3.5 Flash, plus StepAudio 3 Realtime, TTS, Gen, Music and ASR.
What is Step 3.7 Flash?
Step 3.7 Flash is a multimodal Flash model built for agent workflows. It handles office and coding tasks and reads PDFs, tables, charts and screenshots.
Does StepFun provide an API?
Yes. The StepFun API at api.stepfun.ai uses a chat completions format, and models such as step-3.7-flash are available.
Is Step 5 open source?
Step 5 Preview's model weights are available on Hugging Face and GitHub, according to StepFun's site.
Key Takeaways
- StepFun builds Step 5 Preview, Step Flash models and StepAudio 3.
- API and the Flash Pro plan are paid; Step 5 Preview weights are free to download.
- Best for developers; hosted use sends data to a Chinese company.
Best StepFun Alternatives
- DeepSeek - Chinese AI chat assistant and low-cost API with a 1M-token context window.
- Qwen - Alibaba's model family with open-weight models you can download and an API on the Alibaba Cloud platform.
- Zhipu AI - Zhipu AI is a Chinese AI company with GLM models, an agent assistant, a coding plan for developer tools, and an API for sale.
- Claude - Anthropic's assistant, strong at writing, analysis and coding, with an API for developers.
Bottom Line: StepFun is a strong developer-focused AI company, with Step 5 Preview open weights, fast Flash models and the StepAudio 3 voice family. Try it in AI Studio, and self-host when data must stay with you.
Sources checked November 2026: StepFun site.
Last Reviewed: November 2026 | Reviewed by theaitoolsbox.com editorial team
Pricing Plans
Flash Mini
$6.99 per monthEntry Step Plan tier with 400M Credits a month, for coding and agent tools. $18.99 quarterly, $69.99 yearly.
- 400M Credits per month
- Use in Claude Code, Trae, Cursor and OpenClaw
- No time-window or request-frequency limits
Flash Plus
$9.99 per monthStep Plan tier with 1,600M Credits a month, for coding and agent tools. $26.99 quarterly, $95.99 yearly.
- 1,600M Credits per month
- Use in Claude Code, Trae, Cursor and OpenClaw
- No time-window or request-frequency limits
Flash Pro
$29 per monthStep Plan tier with 8,000M Credits a month, for coding and agent tools. $79 quarterly, $289 yearly.
- 8,000M Credits per month
- Use in Claude Code, Trae, Cursor and OpenClaw
- No time-window or request-frequency limits
Flash Max
$99 per monthLargest Step Plan tier with 40,000M Credits a month, for coding and agent tools. $269 quarterly, $989 yearly.
- 40,000M Credits per month
- Use in Claude Code, Trae, Cursor and OpenClaw
- No time-window or request-frequency limits
Free open weights
Free to downloadStep 5 Preview weights are free to download on Hugging Face and GitHub.
- Step 5 Preview weights
- Available on Hugging Face
- Available on GitHub
Prices checked November 22, 2026. Confirm on the vendor site before buying.
View Full Pricing on WebsiteAPI Pricing
step-5-preview
$1.00 input / $2.70 output per 1M tokensMultimodal reasoning model. Cache hit input is $0.10 per 1M tokens. Output tokens include reasoning and the final answer.
step-3.7-flash
$0.20 input / $1.15 output per 1M tokensMultimodal reasoning model tuned for agents. Cache hit input is $0.04 per 1M tokens.
step-3.5-flash
$0.10 input / $0.30 output per 1M tokensReasoning model tuned for agents. Cache hit input is $0.02 per 1M tokens.
StepAudio speech
$0.125 per session minute (realtime)New speech prices take effect 12:00 on November 23, 2026 UTC+8. TTS is $14.00 per 1M input text characters, ASR max is $0.60 per audio hour, voice cloning and design $1.50 per voice.
step-1o-turbo-vision
$0.36 input / $1.15 output per 1M tokensVision model. Cache hit input is $0.07 per 1M tokens. Image input counts as tokens, 169 tokens per image by default.
API rates checked November 22, 2026. Rates can change; confirm on the vendor site.