What is Step 5 Preview?
Step 5 Preview is a sparse mixture-of-experts model with 600B parameters, of which 27B work on each token. It reads text, up to 60 images and video in one request, holds 1M tokens of context and calls tools through StepFun's API, with low, medium and high reasoning settings. StepFun built it for software engineering, long research tasks and finance.
It is StepFun's current flagship and sits above Step 3.7 Flash, which stays the cheaper, faster option. StepFun has promised open weights but has not published them yet, so for now you reach it through StepFun's own API, Step Plan and AI Studio.
How is Step 5 Preview different from Step 3.7 Flash?
Step 5 Preview is the larger, stronger and pricier model with four times the context; Step 3.7 Flash is the cheaper, faster one with weights you can download now.
- Size
- 600B parameters with 27B active, against 198B with about 11B active.
- Context
- 1M tokens, against 256K for Step 3.7 Flash.
- Price
- $1.00 / $2.70 per million tokens, against $0.20 / $1.15.
- Weights
- Step 3.7 Flash is open under Apache 2.0; Step 5 Preview weights are promised but not released.
- StepFun's advice
- Step 5 Preview for software engineering, knowledge work and multi-step agents; Step 3.7 Flash for high-speed multimodal work.
Who is Step 5 Preview best for?
Step 5 Preview is best for finance research, long-document analysis and budget coding agents that need a 1M context and image or video input.
Financial research and valuation
StepFun tuned it for finance, and it scores 66.4% on FrontierFinance against 55.0% for GPT-6 Astra, though still behind Claude Opus 5 at 69.7%.
Research across many long sources
It holds 1M tokens and scores 88.3% on AA-LCR v1.1 in StepFun's table, above Claude Opus 5 at 79.3%.
Coding agents on a budget
StepFun documents a Claude Code setup with the full 1M window, and its Step Plan covers the model from $6.99 a month.
Agents that read screens, charts and video
One request can carry up to 60 images plus video, so an agent can work from screenshots, charts and recordings.
What's included with Step 5 Preview?
Image and video input, three reasoning levels, automatic prompt caching, tool calling with JSON output, setups for common coding tools, and StepFun's own Step Code agent.
- Image and video input
- JPG, PNG, WebP and static GIF images by URL or Base64, up to 60 per request, plus MP4, QuickTime or Matroska video.
- Reasoning effort
- Low, medium or high per request; xhigh and max settings from other tools map to high.
- Automatic prompt caching
- Prompts over 256 tokens are cached on their own, and cache hits cost $0.05 per million tokens, 5% of the normal input rate.
- Tool calling and structured output
- Function calling plus JSON Mode and JSON Schema, through OpenAI-style Chat Completions or Anthropic-style Messages requests.
- Works in common coding tools
- StepFun has setup guides for Claude Code, Codex CLI, Cursor, Cline, Kilo Code, OpenClaw and others.
- Step Code
- StepFun's own terminal coding agent, built for Step models, with a goal mode for tasks that run for hours.
How much does Step 5 Preview cost?
Step 5 Preview costs $1.00 per million input tokens and $2.70 per million output tokens on StepFun's API, or comes with Step Plan from $6.99 a month; there is no free plan.
- Cached input $0.05 per million tokens
- Reasoning tokens are billed as output
- Web search $0.006 a call, image search $0.02
- 400M Step Plan credits a month
- Use in Claude Code, Cursor, OpenClaw and other tools
- Also covers Step 3.7 Flash and Step 3.5 Flash
- 1,600M Step Plan credits a month
- Priority API rates
- Priority technical support
- 8,000M Step Plan credits a month
- Priority API rates
- Priority technical support
Flash Max costs $99 a month for 40,000M credits. Paid quarterly, Mini, Plus, Pro and Max cost $18.99, $26.99, $79 and $269; paid yearly, $69.99, $95.99, $289 and $989. StepFun values about 7M credits at $1 of model use; credits arrive monthly and expire at month end, and booster packs add 400M credits for $6.99 or 1,600M for $9.99. API rate limits rise with total top-ups, from 100 requests a minute. StepFun lists no free plan for this model. Prices are in US dollars. Full pricing on platform.stepfun.ai. We checked these prices on .
How do readers rate Step 5 Preview?
No one has rated Step 5 Preview yet. Be the first.
Sign in to rate. One rating per account.
Step 5 Preview pros and cons
It matches Kimi K3 and GLM-5.3 on many of StepFun's tests at a low price and reads images and video, but it trails Claude Opus 5 and GPT-6 Astra on the hardest agent work and its weights are not out yet.
Pros
- Low price for its class: $1.00 / $2.70 per million tokens, and Hacker News and Reddit users point to the price for performance as the main draw.
- Close to Kimi K3 and GLM-5.3 on StepFun's coding tests, with 67.7% on DeepSWE v1.1 against 67.5% and 66.9%, at a third of Kimi K3's input price.
- Reads images and video as well as text, which GLM-5.3 cannot, across a full 1M-token context.
- Strong on finance and long-document work in StepFun's table, ahead of GPT-6 Astra on FrontierFinance and of Claude Opus 5 on AA-LCR.
Cons
- Well behind the top closed models on the hardest agent work: 33.3% on Terminal-Bench v4 against 52.3% for Claude Opus 5 and 57.9% for GPT-6 Astra, and StepFun says a clear gap remains on its most demanding long tasks.
- The open weights are promised but not out, and at 600B parameters they will need server-class hardware; one Reddit estimate puts a 4-bit copy at about 512 GB of memory.
- Output stops at 64K tokens per response, and every reasoning token is billed at the $2.70 output rate.
- No free plan is listed: AI Studio runs the model after you sign in, but StepFun publishes no price or limits for it there.
What are the best Step 5 Preview alternatives?
Step 3.7 Flash for open weights and a lower price, GLM-5.3 for terminal work, and Kimi K3 for open weights that read images.
GLM-5.3FreeChoose it when terminal and automation work matter most: it beats Step 5 Preview on Terminal-Bench v4 in StepFun's own table and costs $1.40 / $4.40 per million tokens, but reads text only.
Kimi K3FreeChoose it when you want open weights that read images now: it scores higher on MMMU-Pro and SWE-Marathon in StepFun's table, though its API costs $3 / $15 per million tokens.
Questions about Step 5 Preview
What is Step 5 Preview?
Step 5 Preview is StepFun's new flagship model for coding, agents and finance work, a 600B-parameter model that reads text, images and video across 1M tokens for $1.00 / $2.70 per million tokens.
Who is Step 5 Preview for?
Step 5 Preview is best for finance research, long-document analysis and budget coding agents that need a 1M context and image or video input.
Is Step 5 Preview free?
No free plan is listed. StepFun's API charges $1.00 / $2.70 per million tokens and Step Plan starts at $6.99 a month. StepFun's AI Studio runs Step 5 Preview in the browser after you sign in, but StepFun does not publish a price or limits for it.
How much does Step 5 Preview cost?
Step 5 Preview's paid plans are API, pay as you go at $1.00 / $2.70 per million input / output tokens, Step Plan Flash Mini at $6.99 per month, Step Plan Flash Plus at $9.99 per month and Step Plan Flash Pro at $29 per month.
What are the best Step 5 Preview alternatives?
The closest alternatives on Recaply are GLM-5.3 and Kimi K3.
Where can I try Step 5 Preview?
On its official website, stepfun.com.
Is Step 5 Preview open source?
Not yet. StepFun says Step 5 Preview will be released with open weights, but they are not on its Hugging Face page and the licence has not been announced. Until then you use it through StepFun's API, Step Plan or AI Studio.
Can I use Step 5 Preview in Claude Code?
Yes. StepFun's Step Plan gives Claude Code an Anthropic-style base URL and the model name step-5-preview. To use the full 1M window, set CLAUDE_CODE_MAX_CONTEXT_TOKENS to 1000000, or Claude Code assumes 200K.
How big is Step 5 Preview?
It has 600B parameters in a sparse mixture-of-experts design, and 27B of them work on each token. That makes it about three times the size of Step 3.7 Flash, which has 198B parameters with about 11B active.
Our verdict
Step 5 Preview is a good-value model for finance, research and coding agents that need 1M tokens and image input, as long as you do not need top-tier results.
On StepFun's own tests it lands level with Kimi K3 and GLM-5.3 at $1.00 / $2.70 per million tokens, and it reads images and video as well as text.
- Skip it for
- The hardest long-running coding and terminal work, where StepFun's table shows Claude Opus 5 and GPT-6 Astra well ahead.
- Check first
- Run it on your own tasks and count reasoning tokens, since they are billed at the output rate.
Step-3.5-Flash
Qwen3.6-Max-Preview
Step3-VL-10B
Step 3.7 Flash
Qwen3.7-Max