A sparse flagship, sold before it is shared
StepFun launched Step 5 Preview on Sunday, a sparse mixture-of-experts model with 600 billion total parameters of which 27 billion are active per token. It takes a one-million-token context, accepts text and images, and is available through the company’s API and Studio from the day of the announcement. StepFun says the weights will open on 15 October.
The pricing is the headline. Step 5 Preview costs $1.00 per million input tokens and $2.70 per million output tokens, with a 95 per cent discount on cached input. That is aggressive for a model of this size, and it is the number the company clearly wants the market to react to.
The active-parameter count is what makes it possible. A model that routes each token through 27 billion of its 600 billion parameters costs roughly what a 27-billion-parameter dense model costs to serve, while holding far more in memory. The trade is memory footprint for compute per token, and StepFun has taken it as far as most of its peers.
Uneven where it counts
Artificial Analysis scored Step 5 Preview at 44 on its intelligence index, an independent composite. That places it in the upper band of open-weight models without troubling the frontier.
StepFun’s own published figures show the shape of the model more clearly than the composite does. It reports 66.4 on FrontierFinance against 55 for OpenAI’s GPT-6 Astra, and 33.3 on Terminal-Bench v4 against Astra’s 57.9. These are the vendor’s numbers, not independent runs, and they point the same way: the model is strong on structured professional-knowledge work and financial analysis, and well behind the frontier on long-horizon terminal and coding tasks.
That is an awkward combination for a model positioned at agentic work, because agentic coding is exactly where Terminal-Bench measures. It is a better fit for the financial-analysis and knowledge-work use cases StepFun names alongside it.
The open-weight promise
Founder Jiang Daxin is running a split strategy: an open-weight model, Step 3.5 Flash, that developers can deploy and adapt, and a closed flagship the company meters. Step 5 Preview is supposed to cross over on 15 October.
There is reason to hold that date loosely. A Hugging Face repository named stepfun-ai/Step-5-Preview-BF16 existed on the afternoon of the announcement containing a single .gitattributes file — no weights, no configuration, no licence and no model card. An empty placeholder is not a release, and the licence, which determines whether anyone can use the model commercially, has not been published.
StepFun raised $718m in a Series B+ round in January 2026, with Tencent among the backers. The question the next month answers is whether $1 per million input tokens is a price the company can sustain, or an introductory rate — and whether the weights that would let anyone check the cost of serving it actually arrive on the date promised.
