StepFun opens Step 5 Preview API — 600B MoE, 27B active, weights Oct 15

StepFun said Step 5 Preview is available through its API and Studio — a 600-billion-parameter sparse Mixture-of-Experts model that turns on 27 billion parameters per token, with a 1-million-token context and vision input — while full open weights are scheduled for October 15, 2026.

SOFTWARE desk — Frontier labs are splitting “API today” from “weights later.” A 600B MoE that only burns 27B per token is an economics story as much as a benchmark story: if the agent loops hold up, cheap long-context inference is how more teams actually run overnight coding and research agents.

StepFun calls Step 5 Preview its new flagship model for agentic work. Agentic here means the model is meant to keep working through a job — write code, look things up, use tools — instead of answering one chat turn and stopping. The company says it is built for software engineering, professional knowledge work, and finance-style research. File that positioning as StepFun’s. This desk did not sit on an overnight coding loop.

Architecture, still company: a sparse Mixture-of-Experts, or MoE. MoE means the model keeps a large set of specialist parameter groups and only turns on a few for each token it writes. StepFun says the full model has 600 billion parameters and activates 27 billion per token. A parameter is one of the learned numbers inside the model. File the 600B / 27B / sparse-MoE picture as StepFun’s, repeated in same-day coverage. This desk did not count a weight file.

Context and input, company plus Artificial Analysis: a 1-million-token context window, and native text and image input with text out. A context window is how much the model can hold at once — code, filings, prior tool output — in one sitting. One million tokens is enough room for a large repo or a long research pack, not a short chat. File that 1M / vision-in picture as StepFun’s, with Artificial Analysis listing the same window and modalities. This desk did not meter a token counter.

Access split, still company: the API is open now; open weights are scheduled for October 15, 2026. An API, or application programming interface, is a hosted call over the internet. Open weights are the downloaded numbers that would let you run the model on your own machines. Data Studios is explicit that the model is callable on StepFun’s infrastructure today and is not yet an open-weight download. Do not read this card as weights you can pull tonight. File the October 15 date as StepFun’s schedule — not as a completed release.

Independent eval card, Artificial Analysis — not a StepFun scoreboard: Intelligence Index 44; listed API pricing about $1.00 per million input tokens and $2.70 per million output tokens; output speed about 99.8 tokens per second; about $0.71 cost per Intelligence Index task. Artificial Analysis also lists the model as proprietary today, as a reasoning model, and as 600 billion parameters with a 1M context. File those figures as Artificial Analysis’s, and treat listed prices as changeable. This desk did not rerun the index.

Marketing comparisons stay attributed. AIBase, summarizing the launch, says Step 5 Preview pushes a “Pareto frontier” of capability versus cost and that a single-task cost is about one-eighth of Claude Opus 5. Pareto here means a claimed better trade-off — more capability for the same spend, or the same capability for less. File those “Pareto” / “one-eighth Claude” lines as company or secondary claims. They are not independently verified Bad Signal measurements.

Do not confuse this card with Qwen-Image-2.1, a same-day Alibaba image-model release. That is a different SOFTWARE track. This filing is StepFun’s Step 5 Preview API only.

Plain English for the rest of the card: StepFun / 阶跃星辰 = the Shanghai lab that built Step 5 Preview. MoE / Mixture-of-Experts = a large model that only turns on a slice of its parameters for each token. API = a hosted call over the internet. Studio = StepFun’s in-browser try-it surface. open weights = the downloaded numbers for running the model yourself — scheduled October 15, not available today. context window = how much it can hold at once. Intelligence Index = Artificial Analysis’s composite score across coding, knowledge, and agent tasks. agentic = kept on a job across many steps, not one chat reply. This filing is the API-now / weights-later announce plus the AA card — not a weight download, and not a desk bake-off.

PRIMARY here: StepFun’s Step 5 Preview product page — Tier A PRIMARY company source, the original record — plus Artificial Analysis’s independent eval / pricing card for Step 5 Preview, and same-day AIBase and Data Studios coverage of the API-now / weights-October-15 split. The flagship-for-agentic-work line, the software-engineering / knowledge-work / finance positioning, the 600B / 27B sparse MoE, the 1M-token window, text-and-image input, the products-and-API availability, and the October 15 open-weight date are StepFun-attributed. The Intelligence Index 44, the about-$1.00 / $2.70 per million token prices, the 99.8 tokens-per-second speed, and the about-$0.71 cost-per-task figure are Artificial Analysis-attributed. “Pareto frontier” and “one-eighth Claude” stay company or secondary claims — not desk measurements. NOT claimed: that open weights are downloadable today, an independently verified bake-off versus Claude or GPT, that this desk called the API or hosted the 600B pool, a stock tip, or investment advice. Distinct from the already-filed qwen38-omni-flash, claude-cowork-chat-merge, xai-grok-build-memory, and coder-claude-code-agent-relay. Distinct from Qwen-Image-2.1, a same-day Alibaba image-model release.