Step 5 Preview

Eesel AI's practitioner review explains the Step 5 Preview launch from StepFun (阶跃星辰), a Shanghai lab with a history of shipping the "Step" family across text, vision, and audio. The model is framed for complex work and agentic tasks with particular strength in software engineering and finance, and StepFun's framin The review explains how the sparse MoE works: Step 5 holds 600B total parameters in memory but activates only 27B per token via a router that picks a handful of expert sub-networks for each token, which is the whole cost story. The headline spec everyone latched onto is this 600B/27B shape, enabling frontier-sized capa

MarkTechPost covers Step 5 Preview as StepFun's new flagship for agentic work targeting software engineering, professional knowledge work, and finance, with the launch pitched around comparable intelligence at substantially lower task cost (the Pareto frontier framing). The model is deployable as a hosted API and on th The article details the sparse MoE design: about 4.5% of the 600B total parameters activate per token (27B active), which is central to the pricing story. Documented specifications include model ID step-5-preview, 1M-token context, text/image/video input, text output, reasoning effort levels (low, medium, high), stream

AIbase reports that StepFun fully opened Step 5 Preview as its new flagship base model positioned for real-world Agentic tasks, aiming to expand the Pareto frontier of capability, efficiency, and cost. The sparse MoE architecture holds 600B total parameters but activates only 27B at a time, natively supports up to 1 mi The piece adds detail on StepCodeBench, covering 553 independent code repositories, 9 task types, 20 application areas, and 33 programming languages, with Step 5 Preview ranking strong in overall success rate and cross-scenario stability across bug fixing, feature development, refactoring, and environment configuration

CellCog's technical analysis confirms Step 5 Preview launched at 03:15 UTC on September 20, 2026, as an API and product preview, with API pricing of $1.00 per million input tokens on cache miss, $0.05 on cache hit, and $2.70 per million output tokens, with reasoning tokens billed as output. The post documents the model The piece critically observes that Step 5 Preview runs level with Kimi K3 and GLM 5.3 on most benchmark rows and behind GPT-6 Astra and Claude Opus 5 on the hardest coding and agent rows, while noting that the comparison uses StepFun's High mode against the rivals' Max modes, which is the first caveat to understand bef

Aireiter's buyers'-guide piece treats Step 5 Preview as a pilot candidate rather than a default endpoint, noting that StepFun's official account announced it on September 20, 2026, with open weights planned for October 15. The confirmed details include 600B-parameter sparse MoE with 27B active, 1M-token context window, Public token rates reported are $1 input and $2.70 output per million tokens, with the step-5-preview model ID documented; confidence is medium because third-party measurement or listing is needed for verification. The piece flags that operational questions including rate limits, maximum output, function calling, cache

MyClaw's coverage documents Step 5 Preview as a StepFun reasoning model aimed at programming and professional knowledge work, with official documentation and pricing available as of September 20, 2026. Specifications include model ID step-5-preview, 1M-token context window, text/image/video input, text output, tool cal As checked on September 20, 2026, Artificial Analysis reported an Intelligence Index score of 44 and output throughput of 99.8 for Step 5 Preview. The piece notes Step 3.7 Flash has lower token prices and also accepts images and video, positioning Step 5 Preview as the choice for tasks needing more than 256K context or

StepFun introduced Step 5 Preview as its new flagship for agentic work, announcing it on its product page with the framing of advancing the Pareto frontier of intelligence versus cost. The model uses a sparse Mixture-of-Experts architecture with 600B total parameters and 27B active per token, supports a 1M-token contex The announcement positions Step 5 Preview as delivering frontier-level performance at substantially lower task cost than similarly capable models, though no linked technical report is published and open weights are not yet available. The page presents StepCodeBench and other evaluations as evidence of coding strength a