Step 5 Preview

StepFun lists $1.00 input, $2.70 output, and $0.10 cached input per million tokens on its USD pricing page. Reasoning tokens are included in output billing.

On the Chinese pay-as-you-go platform, standard cache-hit input rose to ¥0.70 per million tokens on October 14, 2026 at 00:30 Beijing time. Eligible users who had previously called the model keep ¥0.35. Step Plan is excluded from this change.

Downloading weights is a different deployment choice from paying the official API tariff. A local or rented deployment needs its own compute, storage, networking, and operational budget. The published per-token API row does not quote those costs. The downloadable artifact also has its own license and serving requirements, which should be checked through the official weights source before deployment. Do not treat the absence of an API invoice in a self-hosted setup as evidence of zero operating cost. For a fair comparison, put the complete self-hosting estimate beside the API usage estimate and retain the assumptions used for each. The direct API example uses uncached input, eligible cache hits, and billed output; a self-hosted estimate can use an entirely different cost structure. Neither estimate establishes task quality or response time without separate evidence.

StepFun explicitly includes reasoning tokens in output billing. Counting only the visible final response can leave out part of the output category. Recompute the estimate using the billed usage categories supplied for the account and period. For example, at the published USD tariff, an additional one hundred thousand billed output tokens adds $0.27, regardless of whether the application displayed those tokens as final-answer text. This is an arithmetic illustration, not a claim about how many reasoning tokens a particular task uses. Also verify cache eligibility and the regional schedule: the Chinese pay-as-you-go cache adjustment has an existing-user exception, while the USD page publishes its own tariff. A reconciliation should identify the actual service address, applicable currency, category totals, and date. If any of those are unknown, preserve the uncertainty rather than choosing the cheapest row by assumption.