StepFun Releases Step 5 Preview With 1M Context for Agent Work

StepFun has released Step 5 Preview, a model aimed at software engineering, professional knowledge work, and finance that is available through its hosted platform and API. Its headline specification is a 1M-token context window, intended for work that can involve large collections of material and multiple steps toward a deliverable.

The release is built around a sparse mixture-of-experts design. Step 5 Preview has about 600 billion parameters in total, according to StepFun’s reported specifications, but activates about 27 billion parameters for each token it processes. Those are different measures: the first describes the full model, while the second describes the portion used for a response.

A hosted start, with a later self-hosting promise

For now, StepFun directs developers to its Open Platform, including its Chat Completions and Messages APIs. That makes Step 5 Preview an API product rather than a downloadable model today. MarkTechPost reported that StepFun plans to make open weights available on October 15, which would give developers a separate route to run the model themselves.

The context window does not supply the surrounding work system

StepFun presents the million-token window as useful for analyzing long documents, multiple sources, and conversation history. The model also supports text, image, and video inputs, but it produces text output. Its maximum output is 64,000 tokens, so its advertised strength is taking in a great deal of material rather than returning an equally vast answer.

The agent framing has an important boundary. Step 5 Preview can use tool calls supplied by an integrating application, but the model does not itself access a user’s local environment or external services. Long-running workflows will therefore depend partly on the search, code-execution, document, and other tools wrapped around the model.

What developers can configure

• Choose low, medium, or high reasoning effort, alongside streaming responses and prompt caching.
• Use tool calling and structured output through JSON Mode or JSON Schema.
• Send up to 60 images in one request; video can be supplied by URL, Base64, or a Files API reference.

A cost pitch alongside uneven performance evidence

MarkTechPost reported listed API rates of $1 per million input tokens on cache misses, $0.05 for cached input, and $2.70 per million output tokens. StepFun’s own documentation confirms that billing is based on actual input and output use, though that page directs readers elsewhere for current prices and cache rules.

The performance picture is difficult to compare cleanly. StepFun reported scores of 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench; the same launch coverage said GPT-6 Astra and Claude Opus 5 scored higher on all three. These figures provide reference points, but they do not show that Step 5 Preview leads those rivals in coding overall.

MarkTechPost also reported an Artificial Analysis Intelligence Index score of 44 and measured API output speed of 99.8 tokens per second. That independent measurement broadens the evidence beyond StepFun’s own results, but the published figures cover different evaluations and do not establish a single across-the-board ranking.

The test is whether the surrounding tools keep pace

StepFun says the model can perform multi-step work with application-provided tools and reported a test in which it coordinated 950 web fetches in a single agent action. That company-reported example does not guarantee reliable autonomous work. For teams, the practical question is whether their tools, retries, safeguards, and review process can turn a million-token context into dependable results.