Step 5 Preview: API pricing, access and reasoning costs

Choose the StepFun platform that serves your account, complete its identity verification and create an API key. The Chinese platform documents an OpenAI-compatible base URL of https://api.stepfun.com/v1. Select step-5-preview in a chat completion request. Reasoning effort can be low, medium or high; medium is the documented default recommendation. Begin with a short representative task before sending a long document or video.

Both platforms document verification requirements. The international entry tier allows five concurrent requests, 100 requests per minute and 500,000 tokens per minute for accounts with under $15 in cumulative cash top-ups. Those limits do not establish a country availability list: check account eligibility before planning a rollout.

The two official platforms publish different price schedules. Keep the account, currency and cache rate together; these are not exchange-rate conversions.

Both reasoning tokens and final answer tokens are billed as output. The first cache write uses the uncached input rate. The Chinese personal entry tier requires real-name verification and allows five concurrent requests, 100 requests per minute and 500,000 tokens per minute below ¥100 in cumulative cash top-ups.

For an illustrative request with 100,000 uncached input tokens and 10,000 total output tokens, the token charge is $0.127 internationally or ¥0.90 in China. If 80,000 of those input tokens are actually cache hits, the same example becomes $0.055 or ¥0.368. The totals include the stated usage and reasoning output.

StepFun now publishes FP8, BF16 and NVFP4 checkpoints for this same model, with 600B total parameters and 27B active per token. Use the checkpoints to self-host the model, or keep using its existing API route.

The official examples use four B300 or eight H200 GPUs for FP8, and eight B300 or sixteen H200 GPUs for BF16. Start from the Step 5-supported vLLM or SGLang recipe, including its reasoning and tool parsers. Context, parallelism and KV-cache allocation determine the usable serving capacity.

The weights use StepFun Model License 1.0. Keep that license with a redistribution and read its conditions for your intended use.

Artificial Analysis reports an Intelligence Index score of 44, output speed of 87 tokens per second and a $1.03 cost for its index workload as checked October 10. The speed excludes time to first token. These numbers describe its tested reasoning configuration and workload, not every effort setting or your application.

For your own evaluation, keep the prompt, reasoning effort and allowed tools constant. Record total output usage, latency and whether the result passes the task. Compare a successful task's cost, rather than output-token price alone.

Nous Research corrected Step 5 Preview’s Hermes Index result to 32.83 on October 9. Keep that agent-harness score separate from Artificial Analysis’s Intelligence Index; the two tests measure different workloads.

Use GPT-6.1 Sol as another coding and reasoning candidate. Gemini 4 Argon is relevant to long-context evaluation, but its restricted access matters before you choose it. The model directory provides other documented options.