StepFun announced Step 5 Preview on 20 September: 600 billion parameters in a mixture of experts, of which 27 billion active per token, a one-million-token context, and open access on the same day at around 1 dollar per million tokens on input, 2.70 on output. That price, for those specifications, would have seemed unrealistic six months ago.
We have been documenting the fall in prices for months. Here is a useful marker for measuring the ground covered.
Putting that price in context
One dollar per million input tokens places this model far below the leading Western models, which sit in a range from a few dollars to several tens depending on the segment.
For comparison, we reported that GPT-6 Astra had multiplied its price by 2.5, and that Sonnet 5's price had risen by 50% as of 1 September.
The usual caveat applies: comparing list prices no longer means much. We have shown this repeatedly with the effort level, token splitting and context tiers. What matters is the cost per task completed, measured on your own cases.
The architecture, and what it allows
600 billion parameters in total, 27 billion active per token: this is exactly the trade-off we described in our article on mixture of experts. The knowledge of a very large model for the compute cost of a mid-sized one.
That largely explains the price. Serving 27 billion active parameters costs infinitely less than serving 600.
The usual reminder applies to those considering self-hosting: it is the total size that determines the memory required, not the active parameters. This model is cheap via API and out of reach of an ordinary machine.
This release joins those of the week: an open agent with 744 billion parameters, an omnimodal model with a one-million-token context, and this one. Three very large models in a few days, from different players, two of them released without a communications campaign. Eighteen months ago, each would have led the news for a week. Today, they pile up. This is exactly the threshold of banalisation we are talking about this evening.
What to take away
This price is more useful information than a benchmark score, because it measures something verifiable.
It confirms what we wrote about the valuation of a startup that trains no model: the base layer is becoming commoditised, and value is moving up to what is built on top of it.
For anyone building something, the practical conclusion has not changed and is reinforced: do not build on a model, build on an architecture that lets you switch models. The landscape is moving too fast for today's right choice to still be the right one in six months.