StepFun previews 600B-parameter model for long-horizon AI agents
StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model designed for software engineering, professional knowledge work and finance. It contains about 600B total parameters, with roughly 27B activated for each token. The model accepts text, images and video, produces text, and provides a context window of 1M tokens. Its documented features include low, medium and high reasoning effort, streaming, tool calling, JSON Mode, JSON Schema and prompt caching.
The model is currently available through a hosted API and the StepFun platform. StepFun says open weights are scheduled for October 15, 2026. The company also reports that Step 5 Preview coordinated 950 web fetches in one agent action and achieved more than three times the end-to-end speedup for long-horizon reinforcement learning. Its architecture uses 92 Transformer layers in a narrow-deep arrangement, according to Pandaily.
StepFun lists API prices of $1.00 per 1M input tokens, $0.05 for cached input and $2.70 for output, including reasoning. Artificial Analysis gave the model an Intelligence Index score of 44 and measured 99.8 tokens per second on the API. The evaluator noted that the model generated 160M output tokens in its test, compared with a 92M median for similar models.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.