StepFun released Step 5 Preview on September 20, a 600-billion-parameter sparse mixture-of-experts model positioned as its new flagship for agentic work, with API access open the same day and full weights promised on October 15. The Shanghai-based lab prices the model at $1 per million input tokens and $2.70 per million output, a fraction of what Western frontier models charge.
Specs and benchmarks
Step 5 Preview activates 27 billion parameters per token, roughly 4.5 percent of the total, and supports a 1-million-token context window with native image input and text output. Extended reasoning and tool-based workflows are built in. Artificial Analysis, the third-party evaluation service, scored it 44 on its Intelligence Index, level with Kimi K3 Max and about three points behind GPT-5.6 Sol, which costs seven times more per output token.
On coding, StepFun’s own numbers put the model close to Claude Opus 5 on low- and medium-difficulty tasks, with a wider gap on long-running assignments. On the DeepSWE v1.1 software engineering benchmark, run at temperature 1.0, the model scored 66.4 against Claude Opus 5’s 69.7 and GPT-6 Astra’s 55.0. On DRACO, a harder agentic suite, it posted 83.3, close behind GLM-5.3’s 82.3 but well short of Opus 5’s 87.6.
The company showed a demonstration where the model autonomously optimized H100 GPU kernels for 24 consecutive hours, reaching 508 TFlops after 22 hours, above the 493 TFlops it attributes to Claude Opus 5. In a separate test, the model designed training data that lifted Qwen3-30B’s accuracy on the AIME24 math benchmark from 53.3 to 60 percent using fewer annotated tokens. StepFun also says the model can work directly with programmable hardware during development, using cameras, serial ports, screenshots and simulated mouse input to debug physical setups, a capability most rivals do not claim.
| Model | Intelligence Index | Output price per 1M tokens |
|---|---|---|
| Step 5 Preview | 44 | $2.70 |
| GPT-5.6 Sol | 47 | $20.00 |
| Kimi K3 Max | 44 | not disclosed |
Open weights, with a catch
The Hugging Face repository for the model exists but is empty apart from a .gitattributes file. StepFun says the full weights land on October 15. That timeline matters because Chinese open-weight releases have become a real competitive force: US companies increasingly use open Chinese models because they run cheaper than closed tools from Anthropic, OpenAI and Google. A US government website has even used a Chinese AI search tool that the FBI said copied Anthropic’s product.
StepFun says Step 5 Preview would rank among the top three open-weight models globally if the scores hold. The claim cannot be verified until the weights and license terms are published, since benchmark runs depend on exact configurations and quantization choices.
The price-performance race
StepFun frames the release as a move on the Pareto frontier, the trade-off between intelligence and cost. Its argument is that 27B active parameters per token puts the model in the same compute class as much smaller models, which is why throughput hits about 100 tokens per second on its API. A 95 percent cache discount applies to repeated input tokens, and Artificial Analysis computes a blended rate of $0.51 per million tokens at a typical cache mix.
Chinese labs have been shipping quickly this month. Alibaba pushed Qwen-Image-2.1 to Hugging Face, though with a research-only license instead of the Apache 2.0 used before. Zhipu closed roughly $5 billion in financing, a record for China’s AI sector, and Kimi continues to iterate on K3. StepFun’s release adds another data point to a pattern investors and US labs are watching closely: Chinese models matching Western benchmarks at a fraction of the price, with each release widening the price gap rather than closing the capability gap.
What it means for buyers
For engineering teams, the practical takeaway is that a model scoring 44 at $0.71 per task changes the math on high-volume agentic workloads. A team running millions of tool-calling steps can pay a seventh of what the same work costs on a US frontier model, accepting a small quality gap on the hardest tasks. Artificial Analysis measured task cost for Step 5 Preview at $0.71, against $2 for Kimi K3 Max and roughly $5 for GPT-5.6 Sol at comparable settings.
The risks are the usual ones for early releases: unproven reliability at scale, unclear data practices, and geopolitical exposure. US and Chinese officials discussed AI this weekend in New York ahead of a leaders’ summit, with export controls on advanced chips deliberately kept off the agenda. That truce could hold or could break, and any change in policy would complicate procurement for Western firms betting on Chinese models in production.
October 15 is the date to watch. If the weights ship on schedule with a permissive license, Step 5 Preview becomes one of the strongest models anyone can download and run, and the fourth major Chinese open-weight release of the season. If the license is restrictive, as Qwen-Image-2.1’s turned out to be, the release looks more like a marketing exercise than a genuine open-weight contribution, and developers will stick with Kimi and Qwen. Either way, the direction of travel is hard to miss: the capability floor keeps rising while the price floor keeps falling, and US labs now face competitors willing to compete primarily on cost.
The timing also lands just ahead of Nvidia’s AI Day in Singapore, where open models and infrastructure frameworks top the agenda, giving StepFun’s release a ready-made audience of developers weighing their model mix for the final quarter of the year.