On September 18, the Chinese company StepFun presented a reasoning model with 600 billion parameters and one million context tokens, charged $1 per million input tokens.
StepFun unveiled Step 5 Preview on September 18, 2026 and opened API access in the process. It is a sparse architecture with 600 billion parameters, a context window of one million tokens, that accepts text and images but only produces text.
On Artificial Analysis' synthesis index, the model reaches 44 points, which places it above the average of reasoning models sold in the same price bracket. Billing is $1.00 per million tokens in input and $2.70 in output. In terms of speed, the measurement records 99.8 tokens per second and a delay to first token of 2.96 seconds.
One detail deserves attention for those who pay for consumption: the model was particularly talkative during the evaluation, with 160 million tokens produced across the entire test battery. For the same price, longer reasoning ends up costing more. The weights are not published and only one provider serves the API for the time being.
Read the full article on the original source
Go to the source →