Step 5 Preview

eesel AI's analysis confirms Step 5 Preview is a frontier model for complex and agentic tasks from Shanghai lab StepFun, with a particular lean toward software engineering and finance. The piece independently corroborates the sparse MoE architecture (600B total, 27B active per token) and notes StepFun's framing as adva The review restates the 65% lower cost claim versus GLM-5.3 and Kimi K3 at comparable intelligence and notes StepFun's healthy open-weights footprint on Hugging Face for earlier Step family models. The author treats the cost chart as a vendor pitch rather than verified verdict, evaluating Step 5 Preview from a practica

Flowtivity confirms StepFun released Step 5 Preview on September 20, 2026, priced at $1.00 per million input tokens and $2.70 per million output tokens, for a 600B-parameter sparse MoE model scoring 44 on the Artificial Analysis Intelligence Index. Per AI Weekly, the score matches Kimi K3 Max and arrives at roughly one The analysis details the 92-layer sparse MoE architecture with 27B active parameters per token, positioned for long-horizon agentic workloads such as software engineering, coding agents, and financial analysis requiring large codebases or document sets held in context. Flowtivity contextualizes StepFun as a Shanghai-ba

MarkTechPost covers StepFun's launch of Step 5 Preview, the new flagship for agentic work targeting software engineering, professional knowledge work, and finance. The model is presented as a sparse Mixture-of-Experts design with approximately 600B total parameters and 27B active per token, roughly 4.5% of weights per The piece confirms the model is deployable as a hosted API and on the StepFun platform, with self-hosting pending open weights on October 15, 2026. Supported features include streaming, tool calling, JSON Mode, JSON Schema, and prompt caching. MarkTechPost notes that 600B parameters require approximately 1.2 TB in BF16

Wealth Engine reports StepFun shipped Step 5 Preview on September 20, 2026, as a 600B-parameter sparse mixture-of-experts model with 27B active parameters per token, priced at $1 per million input tokens and $2.70 per million output tokens. The piece confirms text and image input, a 1M-token context window, and Artific The coverage cites Pandaily's reporting on a 92-layer narrow-deep architecture tuned for long-horizon agent runs, and Artificial Analysis measurements of 99.8 output tokens per second with 2.96 seconds to first token. OrcaRouter's check of the Hugging Face repository found only a .gitattributes file at the time, with n

CellCog's detailed breakdown confirms StepFun announced Step 5 Preview at 03:15 UTC on September 20, 2026, with model ID step-5-preview priced and documented. The piece confirms a 600B-parameter sparse mixture-of-experts design with 27B active per token, a 1M-token context window, 64k output tokens, and text, image, an CellCog documents official StepFun API pricing at $1.00 per million input tokens on cache miss, $0.05 on cache hit, and $2.70 per million output tokens, with reasoning tokens counted as output. The analysis notes Step 5 Preview runs level with Kimi K3 and GLM 5.3 on most benchmark rows and behind GPT-6 Astra and Claude

MyClaw confirms Step 5 Preview's official documentation and pricing went live as of September 20, 2026, superseding earlier unannounced coverage. The article documents API model ID step-5-preview, a 1-million-token context window, text/image/video input, text output, tool calling, JSON Mode, JSON Schema, and reasoning The piece restates official API rates at $1 per million uncached input tokens, $0.05 for cached input, and $2.70 for output including reasoning tokens. MyClaw notes the model is callable via API subject to account permissions, with standard API usage and Step Plan having separate billing arrangements, and frames Step 5

ExplainX frames Step 5 Preview as announced September 20, 2026, and analyzes the Pareto-frontier claim against Artificial Analysis methodology rather than raw leaderboard performance. The piece confirms a 600B total / 27B active MoE architecture, a 1-million-token context window with vision support, and a headline clai The analysis interprets StepFun's launch chart plotting Artificial Analysis Intelligence Index against cost per task, placing Step 5 Preview at the boundary between efficient and workhorse bands. Open weights are stated to arrive October 15, 2026, with the hosted API accessible now via platform.stepfun.ai. Comparison p

StepFun (阶跃星辰) officially announced Step 5 Preview as its new flagship model for agentic work, framing it as advancing the Pareto frontier for intelligence-versus-cost trade-offs. The page documents a sparse Mixture-of-Experts architecture with 600B total parameters and 27B active per token, a 1M-token context wind The StepFun product page provides an extensive benchmark comparison across DeepSWE v1.1 (67.7 High), StepCodeBench (49.0 High), ProgramBench (80.5 High), Terminal-Bench v4 (33.3 High), Agents' Last Exam ALE-CLI (29.5 High), GDPval-AA v2.1 (1566 High), FrontierFinance (66.4 High), and DRACO (83.3 High), measured against