Chinese AI lab StepFun has launched Step 5 Preview, a new flagship language model that is drawing attention less for raw benchmark dominance and more for where it lands on the price-performance curve. The model, a 600-billion-parameter sparse mixture-of-experts (MoE) system with a 1 million token context window, is already being called China's cheapest frontier-class model yet by technology press covering the release.
The Specs: Big Model, Big Context
According to Pandaily, which reported the launch, Step 5 Preview is built on a 600B sparse MoE architecture — a design in which only a fraction of the model's parameters activate for any given token, keeping inference costs manageable despite the enormous total parameter count. The architecture family has become the dominant paradigm for frontier-scale models, used by everyone from DeepSeek to Mistral.
The context window is equally notable: 1 million tokens, which Artificial Analysis describes as roughly 1,500 A4 pages of text. That capacity allows the model to take on entire codebases, long legal documents, or extended multi-document analysis in a single pass. The model accepts both text and image inputs and outputs text, and it runs in reasoning mode, generating intermediate thinking before producing final answers.
Perhaps the most closely watched detail for developers: StepFun plans to release the model's open weights on October 15, according to Pandaily. A preview period behind an API followed by open weights is becoming a familiar rhythm for Chinese labs, giving the community time to evaluate the model before it lands on local hardware.
Benchmark Position: Strong, Not Dominant
On the independent Artificial Analysis Intelligence Index, Step 5 Preview scores 44, placing it 24th out of 200 models tracked on the platform — well above the median of 24 for its class, though not at the very top of the frontier. Artificial Analysis characterizes the model as being "amongst the leading models in intelligence and well priced when comparing to other models of similar price."
Speed metrics are solid as well. The model produces output at roughly 99.8 tokens per second on the platform's evaluation, faster than the average of about 70 tokens per second across tracked models.
One caveat worth noting: Step 5 Preview is unusually verbose. Artificial Analysis measured it generating around 160 million tokens across its Intelligence Index evaluation, compared with a median of 92 million for comparable models. Verbosity matters in practice — more generated tokens mean slower perceived responses and higher output-token bills, which can erode some of the headline pricing advantage depending on the task.
The Pricing Story
Pricing is where Step 5 Preview is making its most aggressive pitch. The model costs $1.00 per million input tokens and $2.70 per million output tokens via API, according to Artificial Analysis — against medians of $1.88 and $10.00 respectively for models in its class. The Eastern Herald's headline framed it plainly: China's 600B AI model at $1 per million tokens.
For cost-sensitive workloads — high-volume document processing, batch summarization, agentic systems that burn through millions of tokens per day — a frontier-adjacent model priced below the class median on both input and output is a commercially significant option. The open-weights release planned for October could push effective costs lower still for organizations able to self-host.
A Crowded Chinese Frontier
Step 5 Preview arrives amid an intense stretch of releases from Chinese AI labs, with Alibaba's Qwen family, DeepSeek, Z.ai, and others all shipping new models at a rapid clip. Technology outlets in China have described the current period as a full-sprint foundational model showdown, with labs reportedly working through what would normally be holiday periods to keep pace with rivals.
For StepFun specifically, the bet is that the combination of frontier-adjacent intelligence, a 1M token context window, multimodal input, and aggressive pricing will win developer mindshare ahead of the October open-weights release. The discussion on Hacker News, where the model's launch drew more than a hundred points, suggests the developer community is paying attention.
What to Watch
Three questions will shape Step 5 Preview's reception in the coming weeks. First, how does it perform in real-world coding, reasoning, and agentic benchmarks outside standardized index evaluations? Second, will the verbosity issue be tuned down before the open-weights release, and what will the system requirements for local deployment look like? Third, can StepFun's API infrastructure handle demand at scale — a challenge that has tripped up cheaper entrants before?
The answers will arrive quickly in a market that measures progress in days. For continuing coverage of StepFun, the Chinese AI ecosystem, and the global competition on model pricing, follow the latest AI developments as they happen.
Stay Ahead of AI
Model launches like this are reshaping what AI infrastructure costs. Keep up with every release at AI Buzz Wire, your source for AI news across models, research, and business.
Read more AI news →
