- Step 5 Preview scores 44 on Artificial Analysis Intelligence Index with 600B parameters and 1 million-token context
- The model costs about $0.71 per benchmark task, versus $5.86 for Claude Opus 5
Chinese AI model developer StepFun (阶跃星辰) on September 20 released Step 5 Preview, its latest flagship foundation model, which scored 44 on the Artificial Analysis Intelligence Index, ranking among the top three open-source models globally, according to the company.
The model trails Qwen3.8 Max and Zhipu.ai’s GLM-5.3 on the benchmark, which evaluates reasoning, knowledge, coding and multi-turn agent tasks.
Step 5 Preview uses a sparse mixture-of-experts architecture with 600 billion total parameters but activates only 27 billion at a time.
It supports a 1 million-token context window and scored 67.7% on the DeepSWE v1.1 software-engineering benchmark, behind only GPT-6 Astra and Claude Opus 5, according to Shanghai-based StepFun.
One-eighth the cost
Its main selling point is cost efficiency. Artificial Analysis estimates that completing a single Intelligence Index task with Step 5 Preview costs about $0.71, compared with $5.86 for Claude Opus 5 — roughly one-eighth as much.
The model activates about 4.5% of its total parameters during inference, reducing the computing required for each task.
The launch comes as Chinese open-source models gain ground globally.
Dominating the open-source game
In August, Chinese open-source models occupied all of the top six spots in AI model aggregation platform OpenRouter’s monthly usage rankings, while China accounted for 41% of global open-source model downloads, according to industry data cited by StepFun.
Rather than competing primarily on parameter size, StepFun is positioning Step 5 around the efficiency with which computing resources are converted into model performance.
Step 5 Preview is scheduled to be fully open-sourced on October 15, allowing developers to deploy the model locally and retain control over their data.
Header image credit: StepFun’s official website
