SHANGHAI — The number that matters in StepFun’s announcement today is not in the benchmarks. It is in the pricing table: one dollar per million input tokens, and model weights coming free of charge in three weeks. That combination places a model scoring in the top 25 of the 200 commercially available AI models tracked by Artificial Analysis — the same performance tier as OpenAI and Anthropic’s current flagships — within financial reach of developers, research institutions, and government agencies that do not have Washington’s chip-export subsidies or Silicon Valley’s subscription budget.
The Shanghai-based startup launched Step 5 Preview on Saturday, making its API publicly available at $1 per million input tokens and $2.70 per million output tokens. By comparison, OpenAI’s GPT-4o costs $5 per million input tokens, while Anthropic’s Claude 3.7 Sonnet costs $3.
That gap is not merely a rounding error or a temporary introductory discount. It reflects a structural difference in how Chinese AI companies are designing and deploying large models.
Step 5 Preview is a 600-billion-parameter sparse mixture-of-experts model. The headline parameter count is less important than how many parameters are active during inference: rather than using the entire model for every request, it activates roughly 27 billion parameters per token.
That design preserves broad knowledge and reasoning capacity while reducing the computational cost of inference—allowing the company to price the model aggressively.
On the Artificial Analysis Intelligence Index, which tracks performance across coding, reasoning, mathematics, and instruction-following, Step 5 Preview scores 44 out of 100, placing it 24th among 200 models. On Terminal-Bench 4.0, a software-engineering benchmark that tests an AI’s ability to complete autonomous coding tasks in a real development environment, it scores 33% — against rival Kimi K3’s 13% and broadly competitive with US frontier models on that specific task type. The model carries a one-million-token context window covering text, images, and video.
The benchmarks matter less than the question they raise: how did a startup founded in April 2023 get here this fast?
StepFun’s founding team came from Microsoft Research Asia. Jiang Daxin, the company’s co-founder and chief executive, was a vice president at Microsoft before leaving to build the company in Shanghai alongside co-founders Zhu Yibo and Jiao Binxing. Yin Qi, former chief executive of computer-vision company Megvii, joined as chairman in January 2026, bringing experience scaling AI products to government and enterprise clients across China. The company raised $717 million in a Series B+ in January 2026 and closed a $2.5 billion round in May, pushing its valuation to approximately $10 billion. Tencent, Qiming Venture Partners, and 5Y Capital are among the investors, alongside state-linked funds that Beijing has deployed broadly across its domestic AI sector.
Open weights, scheduled for release on October 15, extend the competitive logic further. An open-weight model is not just a cheaper product. Any organisation — a university in Jakarta, a government ministry in Nairobi, a research lab in Brasília — can download the weights, fine-tune the model on its own data, and run it on infrastructure it controls. No API key. No usage limits. No American cloud provider in the chain of custody. The open-weight release schedule, if it holds, represents the most direct challenge yet to the premium-model pricing structure that US AI companies depend on.
StepFun says the model performs particularly well on software engineering, financial analysis, and professional knowledge tasks. The first two categories are also among the highest-value use cases for enterprise AI spending globally — which explains the aggressive pricing. At $1 per million tokens, an organisation running hundreds of millions of inference calls per month saves meaningfully against current US market leaders.
What Step 5 Preview does not yet tell us is how the model behaves outside the benchmarks where it has been evaluated. Safety characteristics, content policy, refusal behaviour, and multilingual performance in languages beyond Chinese and English are not part of the launch documentation. Whether the October 15 open-weight release arrives on schedule remains to be seen. Model releases in the frontier AI sector have a long pattern of slipping without public explanation.
The API is live today. The benchmarks are real. The price is posted. Developers who have been waiting for a credible non-US alternative to the current frontier models now have one.
Xi Jinping’s offer of an open-source AI community to BRICS nations at the New Delhi summit earlier this month framed AI access as a matter of global equity. Step 5 Preview is the private-sector extension of that proposition. And as Anthropic’s chief executive found when Beijing called his AI-slowdown plan a Cold War playbook, China does not intend to let Washington set the terms of the debate uncontested. Jensen Huang’s forecast of accelerating chip demand, delivered at a summit in Scotland this week, assumes a market structure that StepFun’s pricing table is quietly dismantling.
