StepFun released a 600 billion parameter mixture-of-experts model featuring 27 billion active parameters and a 1 million token context window. The Step 5 Preview model, which succeeds Step 3.7 Flash, recorded a score of 44 on the Artificial Analysis Intelligence Index, matching Kimi K3 (max) and trailing GLM-5.3 (max) and Qwen3.8 Max by one point. It is currently available via first-party API with a pricing structure of $1 per 1M input and $2.70 per 1M output tokens.

The model's cost per intelligence index task is approximately $0.72, 2.8 times lower than the $2.00 and $2.01 per task charged by Kimi K3 and GLM-5.3. Step 5 Preview achieved a 46% score on Humanity's Last Exam, though it lagged on agentic performance with a 1,566 Elo on GDPval-AA versus 1,668 for Qwen3.8 Max. The model reaches 42% accuracy on the AA-Omniscience benchmark but hallucinates on 43% of attempts, and StepFun plans to release the weights on October 15.