Chinese AI startup StepFun opened API access to Step 5 Preview on September 20, a 600-billion-parameter sparse mixture-of-experts model that undercuts frontier pricing by roughly a factor of seven. The model carries 27 billion active parameters per token and a 1-million-token context window. Artificial Analysis, an independent benchmarking service, rates it at 44 on its Intelligence Index, which puts it level with Kimi K3 Max and well behind the leading US models, but the price is what has developers talking: $1 per million input tokens and $2.70 per million output, with a 95 percent discount on cached input.
StepFun says full open weights will follow on October 15. For now the Hugging Face repository ships only a placeholder file, so the API launch is the real event. The release continues a pattern from Chinese labs: ship a competitive API first, publish the weights weeks later, and let the developer ecosystem build on top before Western rivals respond.
What the benchmarks actually show
An Intelligence Index of 44 is solid but not frontier. Kimi K3 Max, from Moonshot AI, sits at the same score, and the top US models rate meaningfully higher on the same scale. StepFun’s own claims are more modest than some rivals’: the company positions Step 5 as a cost-performance play rather than a record-setter. For workloads that do not need the absolute ceiling of capability, including most summarization, classification, code assistance and retrieval-augmented generation, the gap may matter less than the price.
The cache discount is the sharper weapon. Applications that reuse long prompts, including agents, coding assistants and document pipelines, can see effective input costs fall by an order of magnitude when the 95 percent cached rate applies. That pricing structure is aimed directly at the agentic workloads that have made AI spending balloon for startups and enterprises alike.
The open-weights angle
The October 15 weight release could matter more than the API. Open-weight Chinese models have become the default backbone for self-hosted inference: Tencent’s Hy4-preview, a 770-billion-parameter model with a 1-million-token context, was released under Apache 2.0 in late August, and DeepSeek’s open releases reshaped expectations for what free weights can deliver. If Step 5’s weights match the API’s benchmark scores, developers running vLLM or SGLang on their own hardware get a 1-million-token context model at no license cost, which pressures both closed-model vendors and smaller open-source projects.
It also continues the migration of capability out of the US-only tier. Only months ago, a 600-billion-parameter MoE with a million-token context would have been a flagship announcement for an American lab. Now it arrives from Shanghai at commodity pricing, and the response from US labs has been silence rather than a price cut.
The pricing war in context
Step 5’s pricing lands in a market that has been deflating all year. A year ago, million-token context was a premium feature priced accordingly; now Tencent ships it in an Apache 2.0 open release and StepFun charges $1 for the input side of the same window. The compression is driven by architecture as much as competition. Sparse mixture-of-experts models activate only a fraction of their parameters per token, so the 600-billion-parameter count costs far less to serve than a dense model of the same size. StepFun’s 27 billion active parameters per token is in the same range as DeepSeek’s flagship designs, which proved that training and inference costs at this scale could be cut to fractions of dense-model budgets.
The 95 percent cache discount deserves its own note. Prompt caching has become the hidden variable in AI economics. Agents that carry long system prompts and tool definitions on every call spend most of their budget on repeated input, so a steep cached-input discount converts directly into cheaper agents. Vendors that get caching right win the agentic workload market almost by default, and StepFun’s pricing suggests it understands that.
Where StepFun fits in the Chinese field
StepFun is one of China’s so-called AI tigers, the well-funded startups racing the giants Alibaba, Tencent and ByteDance. The company has raised significant venture capital and builds on its own multimodal foundation models. Its position is awkward: it competes with state-backed giants on capital while trying to differentiate on research quality. A credible Step 5 helps it stay in the conversation as consolidation pressure builds, and reports that Chinese models from seven major developers combined generate only about 10 percent of OpenAI and Anthropic’s revenue underline how much of the race is still about distribution rather than raw scores.
For buyers outside China, the calculus is complicated by geopolitics. StepFun’s API is accessible internationally, but enterprise procurement teams weigh data residency, export-control headlines and vendor stability. The open-weights release sidesteps much of that: weights carry no ongoing vendor relationship, which is exactly why the October 15 date may prove more consequential than today’s API.
What to watch
Three things decide whether Step 5 matters beyond the benchmark tweets. First, whether independent evals confirm the Artificial Analysis score on real coding and agentic tasks. Second, whether the October 15 weights arrive complete and without usage restrictions, which has tripped up some earlier Chinese releases. Third, whether US labs respond on price or cede the mid-market. If the answer to all three goes StepFun’s way, the compression of frontier-adjacent pricing continues, and every AI budget in the world gets a little easier to defend.