Step 5 Preview: the API now, the weights in a month, the licence anyone's guess

On 20 September 2026 the Chinese company StepFun (阶跃星辰) unveiled Step 5 Preview, its new flagship model aimed at agentic work, opening API access the same day. The official documentation identifies it as `step-5-preview` and states a context window of 1 million tokens with a maximum output of 64,000. On the multimodal side the model card is explicit: the model “natively supports text, image, and video input”, up to 60 images per request in JPG, PNG, WebP and static GIF, and video files under 128 MB in MP4, QuickTime or Matroska, with a recommended duration of five minutes or less. Then there is streaming output, tool calling, JSON Mode and JSON Schema, prompt caching and three levels of reasoning effort — `low`, `medium`, `high`. Rate limits, the pricing page specifies, “are based on the cumulative top-up amount on your account”: anywhere from 5 to 10,000 concurrent requests depending on how much you have topped up, from tier V0 to V5.

Price is where the new model parts ways with its predecessors. The official price list gives $1.00 per million input tokens on a cache miss, $0.05 on a cache hit and $2.70 on output. For comparison, from that same list, step-3.7-flash costs 0.20/0.04/1.15 and step-3.5-flash 0.10/0.02/0.30: on input, Step 5 Preview works out to roughly five times the previous multimodal model. On architecture, StepFun claims a sparse Mixture-of-Experts with 600 billion total parameters and 27 billion active per token, 4.5% of the total. That figure, however, does not appear in the technical documentation — which gives no parameter count at all — but in the company's announcement on X, repeated consistently by the outlets that carried the news. It was not possible to open the original post (x.com returns HTTP 402), so it remains a company statement relayed by third parties, not something verified at source.

The part about the weights is, for now, entirely in the future tense. StepFun has said the model's full weights will be made available on 15 October 2026 — again a company statement picked up by the press, not confirmed by the documentation. A direct check against the Hugging Face API on 21 September shows that the official `stepfun-ai` organisation hosts no Step-5 repository: the most recent model published there is Step-3.7-Flash, created on 23 May and last updated on 3 June 2026. The public index does contain at least one third-party repository called Step-5-Preview, whose contents have not been verified and which should not be mistaken for an official release. Above all: no licence for the weights has been announced, neither in the documentation nor in the press coverage, and it is not known whether the release will include the base model as well as the aligned one.

In the meantime, one external measurement we were able to verify comes from Artificial Analysis, which gives Step 5 Preview a score of 44 on its Intelligence Index and confirms the price, the context window and the parameter count. It is a single third-party evaluator measuring an unpublished configuration: until the weights are out, nobody else can reproduce it independently. One detail on the Artificial Analysis page deserves attention: the model is dated 18 September there, two days before the announcement, and was already reachable through the API under the tag `step-5-preview-b`. Why it was queryable and measurable before it publicly existed has not been explained. The broader pattern, for that matter, is not new — StepFun is following the road taken this year by Z.ai's GLM-5.3-Flash and by DeepSeek V4.1-Flash: big model, low price, weights promised. Here, though, nearly a month separates the API from the download, and no licence has been stated.

It is the distance between the two dates that interests me. For four weeks the only way to know this model is to pay for it by the token or to trust whoever measured it: “open weights” remains a future commitment, while the present is a closed API. 15 October will not be a date to tick off but one to read — what will matter is not that the files appear, it is what the licence beside them says.

— Olya