StepFun Step 5 Preview

Overview

Step 5 Preview is StepFun's flagship model for long-horizon agentic work, software engineering, finance, and professional knowledge tasks. Announced on September 20, 2026, it uses a sparse mixture-of-experts architecture with 600 billion total parameters and 27 billion activated per token. The hosted model combines a 1-million-token context window with up to 64K output tokens and accepts text, images, and video.

This release is aimed at applications that must plan, use tools, work across large repositories or document sets, and preserve context through extended task trajectories. StepFun's announcement emphasizes software engineering and professional work rather than positioning the model as a simple chat upgrade.

Step 5 Preview is available now through StepFun's API, AI Studio, and Step Plan. The model weights are not open at launch; StepFun says it plans to release them on October 15, 2026.

What's New

A 600B sparse model with 27B active parameters

StepFun scales the flagship to 600 billion total parameters while activating 27 billion for each token. The sparse design aims to provide high capacity without incurring the full inference cost of a dense 600B model on every token.

The official launch materials do not document a full pre-training, supervised-fine-tuning, or reinforcement-learning stack. They instead describe long-horizon evaluations, including GPU-kernel optimization and a post-training data-loop experiment. Architecture size alone does not determine application quality, so teams should evaluate the hosted model on their own prompts and tool harnesses.

1M context with 64K maximum output

The model can accept up to 1 million tokens and return up to 64K tokens. That capacity is relevant for repository-scale code review, long financial or legal documents, research collections, and agents that accumulate many observations and tool results.

Long context still affects latency and cost. Prompt caching is available for repeated prefixes, and StepFun lists a much lower cache-hit input rate than the cache-miss rate.

Multimodal input carried forward from Step 3.7 Flash

Step 5 Preview accepts text, image, and video input and produces text. Step 3.7 Flash already supported native image and video understanding, so multimodal input should not be presented as a Step 5-first capability. The official Step 5 documentation allows up to 60 images per request and lists JPG/JPEG, PNG, WebP, and static GIF formats. Video can be supplied through a URL, Base64 data, or the Files API in MP4, QuickTime, or Matroska formats, subject to documented size and duration guidance.

This expands agent inputs beyond text documents, but it does not make the model an image or video generator. Its documented output modality is text.

Agent-oriented API controls

The model supports streaming, tool calling, JSON Mode, JSON Schema, adjustable reasoning effort, and prompt caching. Reasoning effort can be set to low, medium, or high, giving developers a way to trade latency and token usage against deeper deliberation.

Tool calling is an interface, not bundled external access. StepFun's documentation says search and code execution are supplied by the integrating application, which remains responsible for tool permissions, execution, and result validation.

Performance Benchmarks

StepFun reports the following selected results on its official launch page. These are vendor-published scores and may depend on StepFun's prompts, tools, scaffolding, and evaluation settings.

StepFun says StepCodeBench covers 553 repositories across nine task categories, 20 domains, and 33 programming languages, and reports 49.0 average@4 for Step 5 Preview. The company also reports a 24-hour autonomous GPU-kernel experiment that reached 508 TFLOPS after roughly 22 hours, illustrating sustained iteration rather than a guaranteed result for every coding agent.

Some launch-page results use internal benchmarks or settings that are not directly comparable with other vendors' published numbers. Treat the table as a starting point for workload-specific testing, not an independent ranking.

Compared With Step 3.7 Flash

Step 5 Preview and Step 3.7 Flash serve different cost and quality targets on StepFun's current platform.

The pricing gap is material: Step 5 Preview's cache-miss input is 5x the Step 3.7 Flash rate, while its output rate is about 2.35x higher. Teams should route only the tasks that benefit from the flagship instead of replacing a lower-cost model without measuring quality and total trajectory cost.

Availability & Access

Use the model ID step-5-preview through StepFun's API. The global platform documents the OpenAI-compatible base URL https://api.stepfun.ai/v1, and AI Studio provides browser-based evaluation before integration.

Step Plan also includes Step 5 Preview for supported coding clients such as Claude Code, OpenClaw, and Cursor. Step Plan and standard API usage are separate channels with different Base URLs, quotas, and billing. Both use a valid Step API Key; Step Plan consumes subscription Credits, while standard API usage is billed separately.

StepFun says the model weights will be released on October 15, 2026. That is a future plan, not current open-weight availability; self-hosting claims should wait for an official repository and license.

Pricing & Plans

The official global pricing page lists the following Step 5 Preview rates per 1 million tokens:

For the standard API, the V0 account tier lists 5 concurrent requests, 10 requests per minute, and 5 million tokens per minute. Higher cumulative top-up tiers raise those limits.

Step Plan is a separate subscription path starting at $6.99 per month for Flash Mini, with Plus at $9.99, Pro at $29, and Max at $99. Subscription value depends on the coding client, workload, and credit consumption, so it should not be compared with API token pricing as though the units were interchangeable.

Who Should Upgrade / Who Should Wait

Upgrade or test Step 5 Preview if:

• Your coding agent needs to reason across a large repository, long tool traces, or several related documents.
• You need one model to interpret text, screenshots, diagrams, or video while returning structured text.
• Your workflow benefits from tool calling, JSON Schema, prompt caching, and explicit reasoning controls.
• You work on finance or professional knowledge tasks represented in StepFun's launch evaluations and can validate performance independently.

Wait or keep a lower-cost model if:

• Step 3.7 Flash already meets your quality target and token cost is the main constraint.
• You require open weights or self-hosting before evaluation; the announced October 15 release has not happened yet.
• You need independently reproduced benchmark results before adopting a preview model.
• Your application depends on bundled web search or code execution rather than supplying those tools itself.

Known Limitations

• Benchmark and long-running agent results are published by StepFun and are not substitutes for independent production tests.
• The model is explicitly a preview, so behavior, availability, limits, and documentation may change.
• The 1M context window can increase cost and latency; it does not ensure that every token is used equally well.
• Search, code execution, and other external actions must be provided and governed by the integrating application.
• The output modality is text even when the input contains images or video.
• Model weights are planned for October 15, 2026 and are not officially available at launch.

FAQ

Sources

• Step 5 Preview announcement - release positioning, architecture, benchmarks, experiments, availability, and planned weight release.
• Step 5 Preview documentation - model ID, context, output, modalities, media limits, and API features.
• Step 3.7 Flash documentation - previous-model architecture, context, multimodal support, positioning, and pricing.
• StepFun pricing - current per-token rates and usage tiers.
• Step Plan overview - subscription tiers, coding-client access, and channel separation.