Analysis
StepFun released Step 5 Preview on September 20, a 600-billion-parameter sparse mixture-of-experts model positioned for agentic software engineering and professional knowledge work, according to AI Weekly and CellCog.
The Architecture: Big Total, Small Active
Step 5 Preview has 600 billion total parameters but activates only 27 billion per token -- a sparse mixture-of-experts design that lets the model draw on a much larger overall knowledge and skill pool while keeping the actual compute cost of each inference call closer to a much smaller dense model. This is the same broad architectural family DeepSeek used to disrupt frontier-model pricing over the past two years, and StepFun's adoption of it here is a continuation of that now-standard Chinese-lab playbook rather than a novel approach.
“The pricing war between Chinese and Western labs remains the more consequential long-term story than any single model's benchmark placement.”
Aggressive Pricing, A Genuinely Long Context Window
On StepFun's own API, Step 5 Preview is priced at $1 per million input tokens on a cache miss, just five cents per million on a cache hit, and $2.70 per million output tokens -- pricing that undercuts most Western frontier labs' current API rates. The model also ships with a 1-million-token context window and multimodal input support across text, image and video, positioning it specifically for long-running agentic workflows that combine reasoning, tool use, code execution and research over extended sessions.
Where It Actually Ranks
On StepFun's own DeepSWE v1.1 benchmark at high reasoning, Step 5 Preview scores 67.7, ahead of Kimi K3 Max's 67.5 and GLM-5.3 Max's 66.9, but meaningfully behind GPT-6 Astra Max's 74.1 and Claude Opus 5 Max's 74.0. That places Step 5 Preview solidly in the upper-middle tier of currently available models -- a real, competitive result, but not the frontier-leading performance the 600-billion-parameter headline number might suggest to a casual reader.
Open Weights Coming October 15
StepFun says it will release Step 5 Preview's open weights on October 15, joining DeepSeek and Alibaba's Qwen line -- whose latest Qwen3 model Pulse covered here -- in China's now-established pattern of shipping large open-weight models that Western developers can download, run locally and fine-tune directly. That distribution strategy pressures closed-model API pricing globally regardless of whether StepFun itself builds a large commercial business, since any developer can substitute a free, locally-run alternative once weights are public.
The Numbers In Context
StepFun enters a genuinely crowded field: AI Weekly counts five frontier-adjacent launches in the ten days around Step 5 Preview's release alone, including Claude Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Muse Spark 1.3 and DeepSeek V4.1-Flash. A mid-tier benchmark result with aggressive pricing and open weights is a coherent strategy in a market this saturated -- StepFun isn't trying to win on raw capability, it's trying to win on cost and openness, the same lane DeepSeek carved out before it.
What To Watch
Whether Step 5 Preview's open-weight release on October 15 drives meaningful third-party fine-tuning and deployment, the real test of whether a mid-tier-benchmark model with aggressive pricing can build a genuine developer ecosystem rather than a one-week news cycle. The pricing war between Chinese and Western labs remains the more consequential long-term story than any single model's benchmark placement.