StepFun’s Step 5 Preview puts frontier-model pricing on a much tighter leash

StepFun has released Step 5 Preview, a large mixture-of-experts model offering a one-million-token context window, image input and API pricing of $1 per million input tokens and $2.70 per million output tokens. The practical significance is not merely that another frontier model has arrived. It is that long-context, agent-focused AI is being priced for ordinary use rather than treated like a ceremonial luxury item.

Watch Desk analysis

What happened

Flowtivity reports that Step 5 Preview has about 600 billion total parameters, with 27 billion active at a time. API access is live, while the company has scheduled a full open-weights release for 15 October 2026. The service also offers a reported 95% discount on prompt caching, which could make repeated instructions, documents or tool definitions substantially cheaper when workloads are structured to reuse them.

The report describes the model as aimed at software engineering, research and professional knowledge work. It cites a score of 44 on the Artificial Analysis Intelligence Index and says early testing has shown strong coding and refactoring performance, alongside less even results in factual knowledge and complex document handling.

Why it matters

The headline price is useful, but the real question is cost per completed task. A million-token context window is only a bargain if developers can use it without paying to resend the same material, and prompt caching matters precisely because agent workflows tend to carry a lot of repeated baggage.

The release also gives smaller or less familiar labs another way to compete: not just by claiming the cleverest model, but by offering a more attractive capability-to-cost ratio. That could put pressure on established providers, particularly for coding agents and other workloads where users care about throughput and reliability more than brand recognition.

Our read

This is a meaningful release worth watching, with one important editorial eyebrow raised at the benchmark language. The supplied report gives concrete specifications and prices, but its performance claims remain claims from an early account, not a neutral head-to-head evaluation. Buyers should test their own workloads, especially where factual accuracy or long documents matter.

The useful takeaway is straightforward: StepFun is betting that frontier AI becomes more competitive when the invoice is part of the product. That is a healthier contest than another parade of giant parameter counts, although the spreadsheet will still demand proof.

What to watch

• Open-weights delivery:
  whether the promised 15 October release arrives on schedule and under what licence.
• Real task costs:
  whether prompt caching produces meaningful savings in deployed agent workflows.
• Independent testing:
  how Step 5 Preview performs against established models on coding, factuality and long documents.
• Reliability at scale:
  whether the reported pricing remains attractive once latency, limits and operational overhead are included.

Discussion spark: Should AI buyers prioritise the lowest cost per token, or pay more for models with better reliability and fewer surprises on real workloads?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.