StepFun announced Step 5 Preview on September 20, a 600-billion-parameter sparse mixture-of-experts flagship with 27 billion parameters active per token, a 1-million-token context window and native image input, and opened its API the same day. The Chinese lab, known until now mostly for its image and video models, is now competing directly in the frontier LLM category, and doing it on price.

According to OrcaRouter's breakdown of the announcement, Step 5 Preview scored 44 on the Artificial Analysis Intelligence Index, tying Kimi K3, a model with 2.8 trillion parameters, and placing among the top three open-source models globally. In the Terminal-Bench 4.0 heavy agentic coding test, it reached 33.3%, close to GPT-5.6 Terra and comfortably ahead of GLM-5.3 Flash and Kimi K3's 12.6%.

Frontier-adjacent performance at a fraction of the cost

The economics are the headline. Superpower Daily's coverage notes the model targets software engineering, professional knowledge work and finance, three categories where agentic workloads burn through tokens fast. Step 5 Preview generates at around 100 tokens per second, well above the 65 average for its class, and its per-task cost is reported at roughly an eighth of comparable frontier models. AI REITER's pricing analysis adds a note of caution: the API path is open, but StepFun's documentation does not yet expose a complete production contract or billing specification, so the model is a pilot candidate rather than a default endpoint for now.

The other open question is the weights. The Hugging Face repository for Step 5 Preview exists but contains only a placeholder file, so true self-hosting is not available yet. The API is OpenAI-compatible on the standard chat completions route, with an Anthropic-compatible Messages route and a documented Claude Code integration for Step Plan users.

Why this matters for marketing teams

Agentic models at this price change the math for AI-assisted production. Research summaries, competitive monitoring, localization at scale, QA of campaign assets and first-pass content drafts are all workloads that live or die on cost per task. When a model near the frontier costs an eighth of the usual rate, workflows that were too expensive to automate last quarter become viable this quarter.

There is also a strategic read. StepFun built its reputation on image and video generation, and its models already power parts of the creative pipeline across the industry, including ours. A lab that ships strong visual models and a frontier-class agentic model from the same stack is moving toward an all-in-one creative intelligence offering, and that is exactly the direction brand and agency tooling is heading.

What to watch

Two things will determine whether Step 5 Preview becomes a real contender or a headline: whether the open weights actually ship, and whether the benchmark results hold up in production workloads once independent teams push it hard. Both should be answered within weeks.

At Miela we track these launches because they land directly in our clients' campaigns: every drop in the cost of capable AI lowers the price of producing, testing and localizing creator content at scale. We will be piloting Step 5 Preview in our own workflows and reporting what survives contact with production.