StepFun announced Step 5 Preview on September 20 and opened API access the same day, shipping a 600-billion-parameter sparse mixture-of-experts model with 27 billion parameters active per token and a 1-million-token context window. The Chinese lab priced the model at $1 per million input tokens and $2.70 per million output tokens, with a 95% cache discount, according to benchmark tracker Artificial Analysis.
Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index, level with Kimi K3 Max and roughly a seventh the price of OpenAI’s GPT-5.6 Sol. The model ranks 24th of the 200 models the tracker evaluates, and it measured output speed at 99.8 tokens per second, above the 70.3 tokens per second median, with a time to first token of 2.96 seconds.
What the model actually ships
The spec sheet is aggressive. A 1M-token limit applies on both the input and output side, so a large codebase or a stack of documents can go into a single call without being split. The model accepts text, images and video, with up to 60 images per request, and reasoning effort is selectable per request at low, medium or high on the same model ID, which lets a caller trade cost against depth without switching endpoints.
StepFun aims the model at software engineering, long-document work and professional knowledge tasks, positioning it as an agentic flagship. The API identifier is step-5-preview, served through the company’s developer platform.
From leak to launch in one day
The announcement closed a strange 24 hours. The same model had been benchmarked earlier in the day under a test tag before StepFun published a product page, and trackers noticed the listing before the company confirmed it. By the afternoon the official announcement was out, the API was live, and the benchmark numbers on the unannounced listing matched the launched model without a single figure changing. Preview models now move from rumor to product without the numbers moving at all, which says something about how standardized evaluation has become.
The leak was harmless in this case. It also previewed the main criticism: that shipping an API and shipping a product are different events, and StepFun ran them a few weeks apart.
The open-weights asterisk
There is a catch worth reading closely. A Hugging Face repository named stepfun-ai/Step-5-Preview-BF16 existed by the afternoon of the announcement, but it contains exactly one file, a .gitattributes placeholder. No weights, no license, no model card, no configuration. StepFun’s own materials put the open-weights release on October 15, so the model is callable today and downloadable in about three and a half weeks, and those are two different things.
That gap matters for buyers. Independent verification of capability claims waits on public weights, and until then the benchmark numbers rest on Artificial Analysis’s own runs against the hosted API. The same tracker notes the model is verbose, generating 160 million tokens across its evaluation set against a median of 92 million, so real-world output cost runs higher than the headline input price suggests. Cost per Intelligence Index task came to $0.71, a figure that folds verbosity into the comparison.
Documentation lags the launch
Independent reviewers flagged the gap between announcement and operational readiness. A September 20 review by AIReiter found no dedicated Step 5 pricing or API page on StepFun’s own platform search, leaving rate limits, maximum output, function calling, cache billing and regional availability without first-party citations. The reviewer’s advice was to treat the model as a pilot candidate rather than a production dependency until StepFun publishes the model page and billing rules.
Conflicting catalog data adds to the caution. Some third-party directories list the context at 1M tokens, others at 100K. Developers testing the API should verify the exact route and limits before hard-coding anything, since a preview launch that opens the API and the documentation at different speeds tends to leave stale listings behind.
Where this lands in the price war
Step 5 Preview arrives into a Chinese open-weights field that has been compressing prices all year. The price of $1 per million input tokens for an Intelligence Index of 44 undercuts frontier Western models by a wide margin, and the open-weights release due October 15 would let anyone run the model on their own hardware, at which point the price floor drops further.
The comparison set is dense. Kimi K3 Max matches the score. DeepSeek’s V4 line has been shipping vision-capable variants since August, including a flash tier that beat Anthropic’s Opus 4.8 on two multimodal benchmarks. Z.ai’s next flagship, referred to in community shorthand as GLM-5.5, has no confirmed model card, price or release date, though observers expect open weights under the same custom terms the company has used all year. StepFun’s bet is that shipping now, cheap and multimodal, beats waiting for a perfect benchmark.
The timing also runs against the political current. US President Trump said on September 19 that he will form an AI Force modeled on Space Force and name an AI czar, dismissing safety concerns as a hoax and vowing not to hinder development. Chinese labs, meanwhile, keep shipping at prices that make export controls look like a speed bump rather than a wall. A 600B model at $1 per million tokens is the kind of data point that shapes both the policy debate and enterprise procurement.
Anthropic’s position adds a second layer. The company called for the industry to slow down, filed for an IPO, and now reportedly weighs a new model launch of its own, according to Reuters. Against that backdrop, a Chinese mid-tier lab shipping a 44-point model at commodity prices reads as an argument that capability diffuses faster than any slowdown pact can contain.
For buyers, the practical read is straightforward. If a workload fits agentic coding or long-document analysis and the cost profile matters, Step 5 Preview is worth a measured pilot at current prices. If production stability, documented rate limits and self-hosting are requirements, the October 15 weights release is the date that actually settles the question.