AIREITERStepFun confirms API access, but third-party catalogs—not a complete first-party rate card—provide the clearest public prices. Treat Step 5 Preview as a controlled pilot, not a default production model.
Step 5 Preview API pricing at a glance
Step 5 Preview is listed at $1.00 per 1 million input tokens and $2.70 per 1 million output tokens by Vercel AI Gateway and EmpirioLabs. Artificial Analysis separately reports a $0.51 per 1M blended estimate under a stated workload mix. These are market reference figures, not a numeric rate card published in StepFun’s cited announcement.
| Item | Current reported value | Evidence status |
|---|---|---|
| Input tokens | $1.00 / 1M | Vercel and EmpirioLabs; verify your route |
| Output tokens | $2.70 / 1M | Vercel and EmpirioLabs; verify billing details |
| Cached input | $0.05 / 1M | Provider catalogs; confirm cache rules |
| Optional web search | $0.013 / call | EmpirioLabs only; verify route billing |
| Model status | Preview | Confirmed by provider listings |
| API access | Available now | Confirmed by StepFun |
What StepFun officially confirms
StepFun’s official Step 5 Preview announcement confirms API access now and separately gives October 15, 2026 as the planned open-weight date. It describes a sparse mixture-of-experts model with 600 billion total parameters, 27 billion active parameters, a 1-million-token context window, and text and vision input for agentic work, software engineering, and finance. These are vendor claims; individual routes may not expose every capability or limit.
Vercel provides the gateway compatibility layer for its listed OpenAI Chat Completions, Responses, and Anthropic-style interfaces, using stepfun/step-5-preview. EmpirioLabs documents its own OpenAI-compatible endpoint with step-5-preview, a 1.02M-token context, 131,072 maximum output tokens, reasoning controls, structured output, and optional web search. Treat those as endpoint-specific implementation details.
Step 5 Preview API pricing and billing caveats
At the reported base rates, estimate cost as:
(input tokens / 1,000,000 × $1.00)
+ (output tokens / 1,000,000 × $2.70)A request with 100,000 input tokens and 20,000 output tokens would cost about $0.154 before taxes, retries, gateway markups, or unreported usage.
| Example request | Input | Output | Estimated model-token cost |
|---|---|---|---|
| Short code review | 60K | 4K | $0.0708 |
| Long-context analysis | 100K | 20K | $0.1540 |
| Agent run | 250K | 25K | $0.3175 |
Cached input is reported at $0.05/M, while EmpirioLabs bills reasoning as output and web search at $0.013/call on its endpoint. Confirm all three rules on your route. Because output costs 2.7 times input, long or reasoning-heavy responses can dominate spend.
Is the API usable in production?
The API is officially available, but the public evidence does not yet establish Step 5 Preview as a production-ready default for every workload. Missing evidence includes route-specific rate limits, SLA terms, compatibility guarantees, and complete billing rules.
| Production question | Established publicly | Still needs verification |
|---|---|---|
| Can I call it now? | StepFun says API access is available | Account, region, and route eligibility |
| What does it cost? | Catalogs report $1/$2.70 per 1M | First-party rate card and billing edge cases |
| What is the model ID? | step-5-preview or stepfun/step-5-preview | Exact identifier on your endpoint |
| How large is the context? | About 1M tokens is repeatedly listed | Actual request limit and truncation behavior |
| Does it support tools or JSON? | Catalogs report support | Route-level behavior and error handling |
| Is it stable? | It remains labeled Preview | Compatibility, SLA, and change policy |
Artificial Analysis reports 99.8 tokens per second, 23.01 seconds to first answer token, and a $0.51 per 1M blended estimate under its 7:2:1 cache-hit/input/output mix. It tracks only one provider, so do not treat the blend as a universal price.
Who should use Step 5 Preview now
Step 5 Preview fits teams that can contain preview risk and have a measurable reason to test it.
Use it for bounded, high-context pilots
Long-document analysis, repository-level code review, research synthesis, and internal agent experiments are sensible starting points. They can benefit from a large context window and low reported input pricing while remaining easy to replay against a fallback model.
Use it when cost matters more than ecosystem maturity
The reported $1/$2.70 rates are compelling for large prompts, repeated context, and capped outputs. Set a maximum output budget rather than allowing every request to consume the model’s full reasoning allowance.
Limits of a preview endpoint
Avoid using it as the sole model for payments, account changes, compliance decisions, or irreversible tool calls. The official announcement describes vision input, while catalogs disagree on some route-level capabilities, so test image handling, function-call arguments, structured output, retries, and timeout behavior independently.
A verification checklist before you ship
- Confirm the endpoint and model ID. Send a harmless request using the exact identifier shown in your provider account. Record the returned model name and region.
- Save usage objects and invoices. Compare input, output, cached, and reasoning-token fields. Do not infer billing from a blended estimate.
- Probe context progressively. Test 50K, 100K, and 350K-token inputs before attempting larger prompts. Record latency, truncation, and error behavior.
- Cap output tokens. Output costs $2.70/M at the reported rate, and reasoning may be charged as output on some routes.
- Test tool calls and keep a fallback model. Check malformed-call recovery, timeouts, policy-sensitive tasks, and retry behavior.
Step 5 Preview API FAQ
What is the official Step 5 Preview price?
See the pricing table above: Vercel and EmpirioLabs report $1 per 1M input tokens and $2.70 per 1M output tokens, but verify the rate and billing fields on the route you use.
What model ID should developers test?
The commonly listed IDs are step-5-preview and stepfun/step-5-preview, depending on the provider or gateway. Check the exact identifier in your account rather than hard-coding a catalog value.
Are reasoning tokens and cached tokens billed separately?
EmpirioLabs says reasoning tokens count as output tokens and lists cached input at $0.05/M. Confirm both rules in your own usage records before forecasting cost.