StepFun released Step 5 Preview on 20 September 2026, a 600-billion-parameter model available immediately through its API at $1 per million input tokens and $2.70 per million output tokens. The Shanghai company has promised to publish the weights on 15 October. It has not said under what licence.
What StepFun Released
Step 5 Preview is a sparse mixture-of-experts model, meaning only a fraction of its parameters run on any given token. StepFun puts the totals at 600 billion parameters with 27 billion active per token, an activation rate of about 4.5%.
- Architecture: Sparse mixture of experts, 600B total parameters, 27B active per token.
- Context window: 1 million tokens, as advertised by StepFun.
- Input types: Text and images.
- Access: StepFun’s API and Studio, through an OpenAI-compatible endpoint, from 20 September 2026.
- Stated targets: Software engineering, professional knowledge work and financial analysis.
As of 21 September 2026, that API is the only way to run the model. There is no downloadable checkpoint, and nothing about self-hosting can be planned until the October release lands with terms attached.
What It Costs
Price is the part of this release that is unambiguous. StepFun charges $1.00 per million input tokens and $2.70 per million output tokens, with a 95% discount on cached input.
Those are low numbers for a model of this size. The independent benchmarking firm Artificial Analysis calculates a weighted figure of $0.71 per task on its Intelligence Index, which is where the comparisons other outlets have been making come from. Several reported that Step 5 Preview matches Moonshot AI’s Kimi K3 Max on index score at roughly 65% lower cost per task.
One caveat travels with the cost figure. Artificial Analysis recorded Step 5 Preview producing about 160 million output tokens across its evaluation set against a median of 92 million, so the model is unusually verbose. Output tokens are the expensive ones, and a chatty model erodes part of its own price advantage in practice.
How It Scores on an Independent Benchmark
Artificial Analysis places Step 5 Preview at 44 on its Intelligence Index, ranking 24th of the 200 models it tracks. The same evaluation records output at 99.8 tokens per second and a time to first token of 2.96 seconds.
That puts it in the upper-middle tier rather than at the frontier. For readers tracking the pace of Chinese releases, it lands in the same stretch of the market as DeepSeek’s V4.1 Flash price cuts earlier this month: competitive on cost per unit of capability, not on raw capability.
Where StepFun’s Own Numbers Show It Behind
StepFun published a comparison table alongside the launch, and it is not a clean sweep. The figures below are the company’s own, as reported by RuntimeWire; they have not been independently reproduced.
| Benchmark | Step 5 Preview | Claude Opus 5 | GPT-6 Astra |
|---|---|---|---|
| FrontierFinance | 66.4 | 69.7 | 55.0 |
| DRACO | 83.3 | 87.6 | 76.8 |
| Terminal-Bench v4 | 33.3 | 52.3 | 57.9 |
The Terminal-Bench row is the one to read twice. On that test, which measures work in a command-line environment, StepFun’s own table puts its model well below both Western frontier models, which matters because agent workloads are the use case StepFun is selling.
The coding figures are messier still. Reports of StepFun’s DeepSWE v1.1 result differ, with 67.7% and 56.6% both appearing in coverage of the same launch. Where a vendor’s own numbers are reported inconsistently, the safest reading is to wait for independent evaluation, as happened with the benchmark claims around GPT-6 Astra’s launch this month.
The Open-Weights Date Comes Without a Licence
StepFun has committed to publishing Step 5 Preview’s weights on 15 October 2026, in BF16 format. It has not published the licence those weights will carry, and that omission decides whether the release is useful to anyone building a product.
“Open weights” covers a wide range of terms among Chinese labs, and the differences are commercial, not cosmetic.
| Model | Licence terms |
|---|---|
| DeepSeek V4 | MIT, unrestricted use |
| Qwen3.6 | Apache 2.0 |
| Kimi K3 | Conditions tied to revenue and attribution |
| MiniMax M3 | Separate commercial agreement required |
| Step 5 Preview | Not published |
StepFun has form here: it distributed Step 3.5 Flash as open weights through Hugging Face, so the promise is consistent with what the company has done before. The point is not that the licence will be bad. It is that nobody can evaluate a deployment until it exists.
The Context Window Is Not Settled
StepFun advertises a 1-million-token context window. The AI routing service OrcaRouter reported that third-party configuration data for the model listed a 350,000-token context with a 64,000-token output ceiling, and that StepFun had not reconciled the two.
That is a single-source observation and StepFun has not responded to it publicly. It is worth knowing before anyone designs around the larger number, because a seven-hundred-thousand-token gap is the difference between a model that can hold an entire codebase and one that cannot.
Who StepFun Is
StepFun was founded in Shanghai in 2023 and is counted among the group of Chinese model developers the local press calls the “AI Six Tigers”. Its chief executive, Jiang Daxin, spent 16 years at Microsoft, latterly as a global vice-president and chief scientist, before starting the company.
Its earlier releases include Step-1V, Step-Audio, Step-Video, Step-3 and Step 3.5 Flash, a 196-billion-parameter model with 11 billion active parameters that was published as open weights. Step 5 Preview is the largest model the company has shipped.
Step 5 Preview FAQ
What Is Step 5 Preview?
It is StepFun’s largest language model, released on 20 September 2026. The design is a sparse mixture of experts with 600 billion total parameters, 27 billion of them active for each token, and it accepts both text and images.
How Much Does Step 5 Preview Cost?
StepFun’s API charges $1.00 per million input tokens and $2.70 per million output tokens, with a 95% discount applied to cached input.
When Will the Step 5 Preview Weights Be Released?
StepFun has named 15 October 2026 and said the weights will be in BF16 format. The licence has not been announced, so what may legally be done with them is unknown.
How Good Is Step 5 Preview?
Artificial Analysis scores it 44 on its Intelligence Index, 24th among the 200 models it tracks. StepFun’s own comparison table shows it behind Claude Opus 5 and GPT-6 Astra on Terminal-Bench v4.
Can I Download and Run Step 5 Preview Now?
No. Until 15 October 2026 the model runs only through StepFun’s API and Studio, and no checkpoint is available for self-hosting.
