StepFun's 600B Step 5 Scores 44 at $1 per Million Tokens

Shanghai-based StepFun announced Step 5 Preview on Sept. 20, 2026, and opened API and Studio access to the model the same day. It is a sparse Mixture-of-Experts system with roughly 600 billion total parameters, about 27 billion of which are active for any given token. The benchmarking firm Artificial Analysis, which has since run the model through its own test suite, lists it with a one-million-token context window and support for text and image input. A flash item on KuCoin's news wire headlined the context window as 100,000 tokens, but the summary inside that same item, like every other account, puts it at one million.

The figure drawing attention is the price. Artificial Analysis lists Step 5 Preview at $1.00 per million input tokens and $2.70 per million output tokens, with a 95% discount on cached input. On the same site, OpenAI's GPT-5.6 Sol (max) is listed at $4.00 and $20.00 — meaning StepFun is asking roughly one-seventh as much for output.

On capability, Artificial Analysis scored Step 5 Preview at 43.7 on its Intelligence Index, rounded to 44 on the model's page, against a median of 24 for comparable models. That places it level with Kimi K3 (max) at 43.6 and below GPT-5.6 Sol (max) at 47.0. Running the full index cost $922.84 on Step 5 Preview versus $3,464.84 on GPT-5.6 Sol, or $0.71 against $1.99 per index task. Artificial Analysis also measured output at about 100 tokens per second, above its 70-token median — with one caveat: the model is, in its words, very verbose, producing 160 million tokens across the index run against a 92-million median. Chatty models erode part of the discount, because output is what users pay the most for.

On the architecture, Pandaily reports a 92-layer narrow-deep transformer stack, which StepFun argues gives longer information paths for implicit multi-hop reasoning while an agent is reading tool output. The same report describes Sparse Grouped-Query Attention with block-wise token merging, which the company says cuts indexer and top-k selection cost to about one-eighth of a denser baseline. StepFun aims the model at long-horizon agent work: coding, software engineering, financial analysis and professional knowledge tasks.

What is not settled is the open-weight release. StepFun says the full weights open on Oct. 15, 2026 — a stated plan, not a delivery. Pandaily reported that when the announcement went out, the model's Hugging Face repository carried only a placeholder .gitattributes file. StepFun's own efficiency claims have not been independently reproduced. If the weights do land, a model that scores within four points of the leaders becomes something anyone can download and run.

Disclosure: NewUJ's editorial process uses Anthropic's Claude models.