StepFun's Step 5 Preview targets long-running agent workflows

A preview for work that spans many steps

StepFun's Step 5 Preview is a model aimed at software engineering, research and professional knowledge work. Its current official documentation describes a one-million-token context window and native support for text, images and video. The output is text: accepting a video as input does not make it a video-generation model.

The model identifier is step-5-preview. StepFun lists a maximum output of 64,000 tokens and positions the large context as a way to handle source documents, code and a long interaction history within the same task. A capacity limit describes what can fit into a request, however, not proof that every detail will be retrieved accurately.

The model and the application have different jobs

The documentation explicitly says that tools, search and code execution come from the application integrating the model. The model does not gain access to a user's files or external services merely because it is described as agentic.

That separation determines what a finished product can do. A model may propose a search or a code change, while the surrounding application must expose the appropriate tool, carry out the request and return the result. Permissions, available data and the quality of that integration therefore remain part of the system being evaluated.

Concrete interface capabilities

StepFun documents streaming, tool calling, structured output, prompt caching and configurable reasoning effort. Its examples include troubleshooting with code and logs, cross-document analysis, chart interpretation and video summarization. Those examples indicate supported use cases; they are not measured completion rates.

The image and video interfaces accept URL and Base64 inputs, with a Files API reference available for video. Their separate size and format limits are documented alongside the model's context window.

Media limits and the cost of a long task

The model overview permits up to sixty images per request, including JPEG, PNG, WebP and static GIF files. For video URLs it specifies individual MP4 files below 128 MB and recommends clips under five minutes. It also lists QuickTime and Matroska among supported video formats through the available input methods. These limits make clear why context size alone is not an upload specification.

StepFun's international pricing page lists $1 per million uncached input tokens, $0.05 per million cached input tokens and $2.70 per million output tokens for step-5-preview. The uncached-input charge includes writing new content to the cache. Output billing includes the model's reasoning as well as its final answer, so a short visible response can still represent a longer charged computation.

The provider separates ordinary pay-as-you-go rate limits from Step Plan subscriptions. Individual API accounts require identity verification, and their rate tier depends on cumulative cash top-ups. Teams assessing access should distinguish the underlying model price from the quota and concurrency arrangement attached to their account.

What the preview establishes

The official page verifies a multimodal, long-context API offering and explains how it is intended to fit into an agent application. It does not, on its own, establish an open-weight release or an independently verified ranking against other frontier systems.

The important question for a longer task is whether the system can keep an accurate working state while using tools and checking a final deliverable. Context capacity helps make such workflows possible, but sustained task quality depends on the model and the software around it together. Step 5 Preview should be assessed at that complete-workflow level.