AIBRIEF
StepFun's Chinese open weight model ranks 24th of 200 on the Artificial Analysis Intelligence Index, with output verbosity that can erode its $1.00/$2.70 sticker price.
StepFun, the Chinese AI lab behind Step-3.7-Flash, opened API and Studio access to Step 5 Preview on September 20, 2026. The 600-billion-parameter sparse mixture-of-experts model activates 27 billion parameters per token, supports a 1-million-token context window, and accepts text and image input.
Artificial Analysis's Intelligence Index v4.3.2 records Step 5 Preview at a score of 44, ranking 24th of 200 models: level with Moonshot's Kimi K3, one point behind Z.ai's GLM-5.3 and Anthropic's Claude Opus 5. Cost per Intelligence Index task comes in at $0.71, against a $1.00-per-million input and $2.70-per-million output list price with a 95% cache discount.
Step 5 Preview generated 160 million output tokens during that index run, roughly 74% more than the 92-million median and ranking 65th of 200 on verbosity, so a workload that looks cheap on input can balloon on output.
Open weights are scheduled for October 15. The Hugging Face repository exists but currently holds only a .gitattributes placeholder, with no weights, license, or model card. StepFun's claims of beating Claude Opus 5 on agentic work by 8x and hitting 508 TFLOPS on its MLA kernel are vendor-run and not independently reproduced.
What remains unknown: a separate developer-tooling configuration in third-party router tooling lists a 350,000-token context with a 64,000-token output cap that StepFun's own materials do not mention.