StepFun launches Step 5 Preview, with open weights due on 15 October
The Chinese lab's flagship is a 600B-parameter mixture of experts activating 27B per token, with a million-token context and pricing of $1 per million input tokens.
StepFun introduced Step 5 Preview on Sunday as its new flagship model for agentic work. It claims frontier-level performance across software engineering and professional knowledge work, with particular strength in finance. The company says it is a 600B total, 27B active mixture-of-experts model with a 1M-token context and vision. It is available now through the API, and open weights are to follow on 15 October.
Figures circulating from the model page put input at $1 per million tokens and output at $2.70. Cached input is discounted by 95 percent, and the model runs at about 100 tokens a second. StepFun reports 67.7 on DeepSWE v1.1, against 74.1 for GPT-6 Astra in its own table, and an Artificial Analysis Intelligence Index score of 44.
Those numbers are StepFun's. One account says Artificial Analysis rates it at about the same intelligence as Kimi K3 Max at around 65 percent lower task cost. The company claims a per-task cost one-eighth that of Claude Opus 5. We have not seen an independent Artificial Analysis post on the model, and the material does not include the prices behind the comparison.
It is the company's first major release in almost two years, according to one developer. He says the model appeared quietly on the platform last week before Sunday's announcement. StepFun said through the day that demand was much higher than expected and that it was working on capacity. TestingCatalog noted the release the same morning.