Step 5 Preview

On September 20, 2026, Shanghai-based StepFun released Step 5 Preview, a foundation model with 600B total parameters that scored 44 on the Artificial Analysis Intelligence Index, tying Moonshot AI's Kimi K3, a roughly 2.8T-parameter model about five times its size. The piece notes that StepFun skipped a Step 4 generati The article reports a single-task cost on that evaluation of about $0.71, roughly one-eighth of what Claude Opus 5 costs for the same work, and frames Step 5 Preview as part of a strategic Chinese-lab pivot toward efficiency over raw scale under constrained capital and compute. It also provides company context: StepFun

StepFun has released Step 5 Preview, a 600B-total / 27B-active sparse MoE flagship aimed at software engineering, professional knowledge work, and finance. According to the article, the model id is step-5-preview, it supports a 1M-token context window, accepts text, images, and video as input, and emits text, with reas Step 5 Preview is available now as a hosted API and on the StepFun platform, while self-hosting waits for open weights that StepFun says will land on October 15, 2026. The piece notes that 600B parameters imply roughly 1.2 TB in BF16 before KV cache, meaning practitioners should plan for multi-GPU server hardware once

StepFun released Step 5 Preview at 03:15 UTC on September 20, 2026, with model id step-5-preview documented and priced. It is a 600B-parameter sparse MoE with 27B active per token, a 1M-token context window, 64k output tokens, and text, image, and video input. StepFun targets software engineering, long-horizon agent ta On StepFun's own API the pricing is $1.00 per million input tokens on a cache miss, $0.05 on a cache hit, and $2.70 per million output tokens, with reasoning tokens counted as output. The model lands level with Kimi K3 and GLM 5.3 on most rows of StepFun's table and behind GPT-6 Astra and Claude Opus 5 on the hardest c

StepFun has fully opened Step 5 Preview as its new flagship base model, positioned for real-world agentic tasks and explicitly framed as pushing the cost-performance Pareto frontier. It is a sparse MoE with 600B total parameters and 27B active per token, supports a 1M-token context window, and processes both text and v The model ranks second only to GPT-6 Astra or Claude Opus 5 on the ALE-CLI subset of Agents' Last Exam, the financial evaluation FrontierFinance, and the cross-domain deep research benchmark DRACO. For programming, StepFun built its own StepCodeBench covering 553 independent repositories, 9 task types, 20 application a