StepFun, the artificial intelligence company based in Shanghai, took an important step on Thursday, October 8, 2026, toward getting its most ambitious model into the hands of developers worldwide. That day, Step 5 Preview appeared on OpenRouter, a gateway that gives access to hundreds of models from different makers through a single interface, and on Vercel AI Gateway. For many programmers, that turns switching models into a single line of code.
The model is not entirely new. StepFun introduced it on September 20 on its own platform and programming interface, and describes it as “StepFun’s flagship model for agentic work.” According to the company, it delivers frontier-level performance in software engineering and professional knowledge work, “with particular strength in finance.” This week’s news is its wide distribution through heavily used services.
Technically, Step 5 Preview is a sparse mixture-of-experts model, an architecture that activates only part of the network for each piece of text. It has 600 billion parameters in total, of which 27 billion are active per token. It supports a one-million-token context window, up to 64,000 output tokens and text, image and video input, as well as tool calling, structured output and selectable reasoning effort levels.
Price is one of its selling points. On OpenRouter it costs $1 per million input tokens, $2.70 per million output tokens and $0.05 per million tokens read from cache. For now, the listing shows StepFun itself as the only provider, which means every request ends up on its servers. Within its catalog, the company offers cheaper alternatives, such as Step 3.5 Flash and Step 3.7 Flash, at much lower rates.
The promise drawing the most attention is open weights. StepFun’s launch page states that the model “will be released with open weights on October 15.” However, as of October 9, the company’s official Hugging Face organization showed no Step 5 repository and no defined license. The site AI Tamer warned that copies of the model have been circulating in third-party accounts since September and are not official StepFun releases.
In the tests the company publishes itself, Step 5 Preview sits close to the best open models and behind the large closed models. On DeepSWE v1.1, a software engineering benchmark, it scores 67.7%, compared with 67.5% for Kimi K3 and 66.9% for GLM-5.3, while GPT-6 Astra and Claude Opus 5 are around 74%. On FrontierFinance, an external evaluation of investment cases, it scores 66.4%, above GPT-6 Astra’s 55% and below Claude Opus 5’s 69.7%.
StepFun acknowledges its limits. On StepCodeBench, an internal test spanning 553 repositories and 33 programming languages, Step 5 Preview scores 49%, compared with 63.9% for Claude Opus 5. The company admits that on the hardest long-running tasks “a meaningful gap to the frontier remains.” On Terminal-Bench v4 it scores 33.3%, below GLM-5.3’s 41.9%. It is worth remembering that the company used the high effort level for its model and the maximum level for rivals, so the comparisons are not equal in compute.
The company also highlights its stamina on long tasks. It says that, given 24 hours to optimize a computing kernel on an NVIDIA H100 GPU, Step 5 Preview reached a peak of 508 teraflops, compared with 493 for Claude Opus 5. In another experiment, it raised a base model’s score on the AIME24 math exam from 53.3% to 60% through an automated post-training process. It even had the model play Pokémon Red, where it passed 3,000 turns and earned three Gym Badges without game-specific tuning.
An independent evaluation by Artificial Analysis gives it 44 points on its intelligence index, compared with a median of 26 in its comparison group, at a weighted cost of $1.03 per task. But the site Magica points out a caveat: the model generated 160 million output tokens in those evaluations, nearly double the median, which suggests it is quite verbose. For a buyer, what matters is not only the price per token but how many tokens it needs to finish the job.
StepFun has significant financial backing. According to January reports cited by Magica, the company closed a B+ round of more than 5 billion yuan, about $718 million, with participation from Tencent, Qiming Venture Partners and 5Y Capital. Its strategy combines foundation models with integrating artificial intelligence into devices.
The launch is part of fierce competition among Chinese labs in open models, where Moonshot AI’s Kimi, Z.ai’s GLM and Alibaba’s Qwen also compete. These models have earned a place in many companies thanks to their balance of quality and cost and the option to run them on in-house infrastructure. This same week, the European search engine Ecosia announced it is replacing Mistral with open models such as Qwen, GLM and Kimi.
For developers, analysts advise caution. Step 5 Preview is, as its name says, a preview, so its behavior and price may change. Before routing real traffic to it, it makes sense to test it on your own long-context and tool-use tasks, compare the cost per completed task and review the data-handling terms of each access route, since StepFun, OpenRouter and Vercel each publish their own.
The key date is October 15. If StepFun then publishes the weights with a clear license, Step 5 Preview will be able to run outside its servers and join the list of large open models competing with closed systems from the United States. Until then, the prudent approach is to treat it as what it is today: a model reachable over the internet, promising and cheap, whose performance figures come mostly from its own maker.
