On September 20, 2026, Chinese AI lab StepFun opened API access to Step 5 Preview, its flagship large language model. The headline numbers are hard to ignore: 600 billion total parameters, 27 billion active per token, a one-million-token context window, and native vision input — all priced at $1 per million input tokens.
The model uses a sparse mixture-of-experts (MoE) architecture, meaning only 4.5% of the network activates for any given token. That efficiency is the same trick behind DeepSeek and Mixtral, and it's what makes operating a 600B model at API scale economically viable. Rather than widening the network, StepFun chose a 92-layer narrow-deep Transformer layout, trading breadth for depth.
StepFun's stated use cases for Step 5 Preview are agentic coding, professional knowledge work, and financial analysis — tasks that benefit from long context and structured reasoning. On the Artificial Analysis Intelligence Index, the model scores 44, placing it in the upper tier of available models while still trailing the current frontier leaders.
What Actually Makes This Launch Different
The open-weights promise is the headline. StepFun says full model weights will be released on October 15, roughly six weeks after the API went live. If that holds, Step 5 would be one of the largest openly available models ever — well above the 70B Llama 3 threshold that most organizations treat as the open-source frontier. For teams who want to self-host or fine-tune at this scale, October 15 is worth blocking in the calendar.
Pricing is designed to compete on cost: $1.00 per million input tokens, $2.70 per million output tokens, and $0.05 per million cached tokens — a 95% cache discount that rewards workloads with stable, reusable prompts. The 1M context window supports long document analysis, extended code repositories, and multi-turn agent sessions without hitting truncation limits.
Architecture and Context
The 1M-token context window is paired with native image input, making Step 5 Preview a multimodal model from launch. That puts it in the same class as GPT-6 and Gemini Ultra in terms of modality support, though the vision capabilities haven't been benchmarked independently yet.
StepFun was founded in 2023 and has been systematically building toward a frontier-class model. Step 5 Preview is their most ambitious release. The "Preview" label is honest about where things stand — open weights and community benchmarks on October 15 will determine whether the technical claims hold under independent scrutiny.
The API is live now at platform.stepfun.com.