Step 5 Preview API Guide: OpenRouter Setup, Costs and Agent Checks

Step 5 Preview is available through OpenRouter for developers who want to evaluate StepFun's model in a coding or agent workflow. Start with its exact API identifier and a small request before connecting an agent that can read files, call tools and accumulate a long conversation. That sequence makes both failures and spending easier to understand.

The details below were checked on October 10, 2026 against the OpenRouter model page and its model catalog API. This is an integration guide, not a report of an agent benchmark we ran.

Step 5 Preview API identifiers and current pricing

The OpenRouter identifier is stepfun/step-5-preview. A provider's native identifier or an agent's display name can differ, so keep the exact gateway name in the configuration. The reviewed catalog lists text, image and video inputs with text output; that inventory does not establish that every client exposes every input mode.

These are the gateway's listed rates on the review date. They are not a subscription entitlement or a promise about a temporary promotion elsewhere. Check the account's actual usage record before using them in a production budget.

The catalog also marks reasoning as mandatory and lists low, medium and high effort levels. Do not assume that disabling a visible reasoning panel disables the model's reasoning work. The display and the billed request are separate things.

Send one controlled OpenRouter request

Use an OpenRouter key supplied through an environment variable. The following request is intentionally small and does not authorize any repository actions:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "model":"stepfun/step-5-preview",
    "messages":[{"role":"user","content":"Explain why an empty list needs a separate case in a mean function."}],
    "max_tokens":1024
  }'

The standard gateway request format is documented in OpenRouter's API overview. Save the HTTP status, returned model name, finish reason and usage values. An output cap can truncate a response that spends part of its budget reasoning; inspect the finish reason before diagnosing a model-quality problem.

If authentication fails, resolve that before changing prompts. If the model is unavailable to the selected route, inspect the provider selection and current catalog. Repeatedly submitting the same request through an agent can turn a straightforward configuration error into a confusing sequence of retries.

Connect the agent after the baseline works

Configure the client to use the gateway base URL and the exact Step 5 identifier. Then test one tool whose behavior is easy to observe, such as reading a small source file. Keep write permissions and automated commands out of this first compatibility check.

A model advertising tool support still needs the client's tool definitions, response parser and execution loop to agree. Record the arguments proposed by the model and the result returned by the executor. For example, a valid JSON argument that points to the wrong file is a task error even if the API request succeeds.

For a coding pilot, use a disposable copy of a repository with a known failing test. Ask for a focused patch and require the agent to show the test result. Review whether it actually ran the test, whether the patch addresses the cause and whether unrelated files changed. A plausible explanation should not substitute for that evidence.

Evaluate context and reasoning budgets

A million-token window describes capacity, not the amount of context every task needs. Start with the relevant files, failing output and a concise statement of the desired behavior. Add surrounding code only when the model needs it to understand the dependency.

Compare a short task and a longer repository task separately. For each, record input size, time to the first useful response, total completion time and whether the outcome passes the same checks. A long history can increase processing time even when the final answer is brief.

After establishing a baseline, compare the supported reasoning effort settings using the gateway's current parameter documentation. Keep the task, tools and output cap unchanged. The useful question is whether additional reasoning improves accepted patches enough to justify the time and tokens it consumes.

Our guide to reasoning-model latency explains why answer length alone is a poor latency estimate. Use that distinction when an agent appears to pause before producing a visible answer.

Measure cost per completed coding task

For an illustrative uncached request containing 10,000 input tokens and 2,000 billed output tokens, the reviewed rates imply $0.01 of input and $0.0054 of output, or $0.0154 before other charges. This is arithmetic from listed rates, not an observed session invoice.

A real agent session may include repeated file reads, tool results, failed attempts and a growing history. Add every request to the cost ledger. Apply the cache-read rate only to the portion the provider reports as cached; repeated text is not proof of a cache hit.

Compare total spend across accepted tasks rather than choosing the cheapest single response. A model that needs another repair cycle can consume more money and attention than a higher-priced request that finishes correctly. Keep the same acceptance criteria when comparing Step 5 with an existing Chinese coding-model route.

Conclusion: adopt Step 5 after the workflow passes

Step 5 Preview API evaluation is most useful when it starts with a controlled request and ends with a reproducible coding task. Verify the model ID, record usage and inspect actual tool behavior. Move more work to the route only after the complete agent loop produces acceptable results within your budget.

Frequently asked questions

Is Step 5 Preview permanently free?

The reviewed OpenRouter catalog lists paid token rates. Promotions on another platform have their own eligibility and expiry; they should not be treated as permanent gateway pricing.

Can every agent use the full context window?

The model's served limit and the client's limits can differ. Check both, and test the amount of context your workload actually needs.

Does a successful API response prove agent compatibility?

No. It establishes a working model call. Tool parsing, file operations, retries and test verification need separate checks.