Step 5 Preview
Step 5 Preview is StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 600B total parameters (27B active per token), a 1M-token context window, and vision input. It targets software engineering, long-horizon agentic tasks, professional knowledge work, and finance, and is available through StepFun's products and API, with open weights planned for October 15.
Key facts
What is Step 5 Preview?
Frontier-level agentic and coding performance at a substantially lower task cost than similarly capable models.
Who is Step 5 Preview best for?
Developers, software engineering teams, AI agent builders, financial and professional knowledge workers, enterprises
What are its main limitations?
- Trails GPT-6 Astra and Claude Opus 5 on several coding and agentic benchmarks
- A meaningful gap to the frontier remains on the most challenging long-running tasks
- Open weights not yet available at launch
- No published pricing on the site (the pricing page returns a 404)
Limitations are based on product materials or editorial synthesis; confirm on the official site.
When was this information verified?
October 17, 2026
Sources
- Official homepageofficial homepage · Oct 17, 2026
Key features
- ✓Sparse Mixture-of-Experts architecture with 600B total parameters and 27B active per token
- ✓1M-token context window with vision input support
- ✓Strong software engineering performance across 33 programming languages and 9 task categories (StepCodeBench, 553 repositories)
- ✓Long-horizon autonomous execution, including 24-hour GPU kernel optimization and multi-thousand-turn agent runs
- ✓Artifact generation for professional work: engineering diagrams, stage plots, moodboards, multi-page reports, and video edits
- ✓Research at scale, coordinating 950 web fetches in a single agent action to assemble 300,000 monthly climate records
- ✓Programmable hardware interaction using cameras, COM ports, screenshots, and simulated mouse input
- ✓Open weights release planned for October 15
Use cases
- →Bug repair, refactoring, feature modification, and CI/CD operations in real codebases
- →Financial analysis and reporting (FrontierFinance, DRA CO evaluations)
- →Long-running optimization loops such as MLA GPU kernel tuning and post-training data improvement
- →Producing finished professional deliverables in domain-specific formats (mechanical, process, manufacturing, creative)
- →Frontend development, 3D asset creation in Blender, and interactive web apps with Three.js
- →Large-scale research and data gathering across many locations and sources
Pros
- +Frontier-level coding scores, including 80.5 on ProgramBench and 67.7 on DeepSWE v1.1
- +Substantially lower task cost at comparable intelligence (Artificial Analysis Intelligence Index score of 44)
- +1M-token context window plus vision input
- +Sustains very long-horizon agentic work with execution feedback
- +Open weights planned for October 15
Cons
- −Trails GPT-6 Astra and Claude Opus 5 on several coding and agentic benchmarks
- −A meaningful gap to the frontier remains on the most challenging long-running tasks
- −Open weights not yet available at launch
- −No published pricing on the site (the pricing page returns a 404)
Positioning
- Core value: Frontier-level agentic and coding performance at a substantially lower task cost than similarly capable models.
- Ideal for: Developers, software engineering teams, AI agent builders, financial and professional knowledge workers, enterprises
- Product type: Available through StepFun products (AI Studio, API Platform) and API
Frequently asked questions
What is Step 5 Preview?
Step 5 Preview is StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 600B total parameters, 27B active per token, a 1M-token context window, and vision input.
How well does Step 5 Preview perform on coding benchmarks?
It scores 67.7 on DeepSWE v1.1, 49.0 avg@4 on StepCodeBench, 80.5 on ProgramBench, 33.3 on Terminal-Bench v4, 29.5 on Agents' Last Exam (ALE-CLI), and 1566 on GDPval-AA v2.1.
How much does Step 5 Preview cost?
No pricing information is provided on the website, and the pricing page currently returns a 404 error.
How can I access Step 5 Preview?
Step 5 Preview is available today through StepFun's products and API, and will be released with open weights on October 15.
What is Step 5 Preview particularly strong at?
It performs strongly across software engineering, agentic tasks, professional knowledge work, and finance, with particular strength in finance.
You may also like
Curated suggestions based on similar categories.
Nemotron 3.5 Super VL
Nemotron 3.5 Super VL is a member of NVIDIA's Nemotron 3 family of open models with open weights, training data, and recipes. It combines visual understanding, reasoning, and tool use in a single model and is described as delivering highest in-class efficiency and accuracy for specialized domains. It can be deployed from local infrastructure to the cloud and is accessible via OpenRouter, NVIDIA NIM API, and Hugging Face.
AIAgentSkills
AIAgentSkills is a curated, community-driven directory for discovering AI agent skills (Claude skills) for Claude, Codex, and other AI agents. It publishes 8K+ searchable skills spanning categories such as writing, coding, design, marketing, testing, DevOps, security, and MCP integrations, each with a short description and install counts.
Cutsio
Cutsio is an AI-powered video MAM (media asset management) platform that lets video teams search every frame of their library by people, dialogue, objects and scenes, then upload, organize, review and deliver from one workspace. It combines visual search, transcript search and agentic chat with share links, collections, upload requests and Drive imports.
Swifly
Swiftly is a Transit Data Platform used by public transit agencies to turn transit data into accurate, actionable insights. It combines AI-powered CAD/AVL, real-time passenger information, proactive operations tools, and reporting and insights so agencies can plan, operate, and communicate with confidence.
raisehunt
RaiseHunt is an AI-powered fundraising platform that combines a 135,152-profile investor database with AI investor matching, AI pitch deck review, personalized automated outreach from the founder's own mailbox, and a built-in fundraising CRM. Founders can paste their startup website to get AI-ranked investor matches and then manage outreach and pipeline in one place.
Vizify
Vizify is an AI agent workspace that hands back finished work — diagrams, infographics, images, videos, Markdown docs, pages, and one-pagers — instead of drafts. Users pick from frontier models across OpenAI, Anthropic, Google, Meta, and xAI, and the agent remembers their formats, review rules, preferred models, and handoff steps for reuse.