StepFun shipped Step 5 Preview on September 20, 2026: a 600-billion-parameter sparse mixture-of-experts model with 27B active parameters, priced at $1 per million input tokens and $2.70 per million output. Artificial Analysis scored it 44 on its Intelligence Index, against a 24 median for comparable models. Open weights land October 15. That input price is 80% below Claude Opus 5.
The model is callable today through StepFun’s API. The weights are not. That gap — a live endpoint now, a Hugging Face repository that currently holds nothing but a .gitattributes file — is the whole story, and it is a pricing story before it is a capability story.
What is StepFun Step 5 Preview?
StepFun Step 5 Preview is a 600B-parameter sparse MoE reasoning model that activates roughly 27B parameters per token. It accepts text and images, claims a 1M-token context window, and is marketed at agentic software engineering and professional knowledge work. StepFun says full open weights arrive October 15, 2026.
The company framed it as a frontier play. “Step 5 Preview is our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance,” StepFun wrote in its launch post on X, under the header “Advancing the Pareto Frontier.”
The architecture, and why it is cheap
Sparse MoE is the entire cost argument. A 600B model that routes only 27B parameters per token pays inference compute on 4.5% of its weight count while retaining the knowledge capacity of the full stack.
Pandaily reported the model uses a 92-layer narrow-deep stack tuned for long-horizon agent runs. That is the same structural bet behind DeepSeek V4.1 Flash’s 552B open-weights drop earlier this month, and behind GLM-5.3-FlashX running at 200 tokens per second on domestic Chinese silicon.
What actually shipped today
- Live: the paid API endpoint, text and image input, extended reasoning.
- Measured: 99.8 output tokens per second and 2.96 seconds to first token, per Artificial Analysis — against a 70 tokens/s and 3.70-second median.
- Promised: open weights on October 15, 2026.
- Missing: the license. OrcaRouter checked the Hugging Face repo and found “no weights, no license, no model card and no configuration.”
How much does Step 5 Preview cost?
One dollar per million input tokens and $2.70 per million output tokens, with a 95% discount on cached input. Artificial Analysis puts the medians for comparable models at $1.88 input and $10.00 output. Against the two US frontier endpoints most enterprises actually bill against, the gap is far wider.
| Model | Input $/M | Output $/M | Context | Step 5 discount (output) |
|---|---|---|---|---|
| Step 5 Preview | $1.00 | $2.70 | 1M (claimed) | — |
| GPT-5.6 Sol | $4.00 | $20.00 | 1.1M | 86.5% |
| Claude Opus 5 | $5.00 | $25.00 | 1M | 89.2% |
| Comparable-model median | $1.88 | $10.00 | — | 73.0% |
On input, Step 5 Preview lands 75% under GPT-5.6 Sol and 80% under Claude Opus 5. Cached, at an effective $0.05 per million, it sits 87.5% under Sol’s $0.40 cached rate.
The verbosity tax nobody prices in
Here is the number that spoils the headline. Artificial Analysis measured 160M output tokens from Step 5 Preview across its evaluation suite, against a 92M median. The model is 74% more verbose than its peer group.
Per-token price is not cost per task. Normalize for that verbosity and the effective output rate is roughly $4.70 per median-equivalent million — still 77% under Sol, but not the 86.5% the price card advertises.
Buyers who learned this lesson with Sakana’s $2-per-million Fugu Max already know to benchmark on completed tasks, not token invoices.
Is Step 5 Preview better than GPT-6 Astra and Claude Opus 5?
Not across the board. Step 5 Preview wins decisively on finance and loses badly on terminal-driven agent work. StepFun’s own published comparisons show it ahead of GPT-6 Astra on FrontierFinance and well behind Claude Opus 5 on Terminal-Bench v4 — which is the benchmark that matters most for the agentic use case it is being sold into.
Where it wins
- FrontierFinance: 66.4 for Step 5 Preview versus 55 for GPT-6 Astra, per StepFun’s reported figures.
- Intelligence Index: 44, versus a 24 median across comparable models.
- Throughput: 99.8 tokens/s, above the 70 tokens/s median.
- Price: the cheapest entry anywhere near that Intelligence Index score.
Where it loses
Terminal-Bench v4 is the problem: 33.3 for Step 5 Preview against 52.3 for Claude Opus 5, a 19-point gap on exactly the long-horizon agent tasks StepFun is targeting. Eastern Herald noted the same score still beats Kimi K3’s 13 — a low bar.
Ranking depends on the cut. OrcaRouter placed the model 24th of 200; RuntimeWire cited 27th of 653. Either way it is a strong debut and not a frontier seat.
The skeptical read: a company claiming “frontier-level performance” on agentic work shipped a model that completes roughly two-thirds as many terminal tasks as the incumbent it is priced against. Cheap tokens do not finish jobs. Tokens that finish jobs do.
The context-window question
StepFun advertises 1M tokens. OrcaRouter found separate documentation indicating 350,000 tokens of context with a 64,000-token output ceiling. Those are not rounding differences — they are different products for anyone building long-running agents, and the discrepancy is unresolved as of publication.
Who is StepFun, and who funded this?
StepFun is a Shanghai-based foundation-model company founded in April 2023 by Jiang Daxin, a 16-year Microsoft veteran who ran Microsoft Research Asia work before leaving. It runs a dual-track strategy: open weights for local deployment, a proprietary API for frontier claims.
RuntimeWire reported a $718M raise in January 2026 for foundation-model development and device integration. Eastern Herald reported a larger figure — roughly $3.2B across two 2026 rounds at a ~$10B valuation, with Tencent and state-linked funds participating. Those accounts do not reconcile, and neither is confirmed by StepFun. Treat the valuation as unverified.
The open-weights track already carries Step 3.5 Flash at 196B parameters. Step 5 Preview is the first time the company has put its flagship on a public release date.
Who wins and loses financially?
The winners are inference buyers with high token volume and tolerance for benchmark variance. The losers are the mid-tier API vendors priced between $2 and $8 per million input tokens who sell on cost rather than capability. Step 5 Preview does not threaten the frontier. It compresses the middle.
Winners
- Agent startups burning tokens on retrieval and summarization. Work that does not need Terminal-Bench-grade reliability just got 80% cheaper.
- Financial-services teams. A 66.4 FrontierFinance score at $1 input is the most defensible claim in the launch.
- Inference hosts. October 15 turns a 600B model into rentable capacity for every neocloud with the VRAM.
Losers
- Anyone selling a mid-tier proprietary API on price. The floor keeps moving, and it keeps moving from Chinese labs — a pattern visible in Cognition’s decision to build SWE-2 on a Chinese base model at 64% less cost.
- US labs defending premium output pricing. Opus 5’s $25 per million output looks increasingly like a reliability premium that has to be re-justified every quarter.
- StepFun itself, potentially. Giving away a 600B flagship six weeks after the API launch is a customer-acquisition bet, not a revenue plan.
What happens on October 15?
StepFun has committed to releasing full open weights for a 600B model on October 15, 2026. If it lands with a permissive license, it becomes the largest recent open-weights drop with a credible Intelligence Index score. If the license is restrictive, the announcement functions mainly as free marketing for a paid API.
Alibaba demonstrated the second outcome this same week: the Qwen team shipped Qwen-Image-2.1 and moved it from Apache 2.0 to a research-only license requiring separate commercial agreements. The word “open” is doing less work in Chinese model releases than it did a year ago.
There is also a policy overhang. Washington’s distillation advisory naming six Chinese firms over token theft means any US enterprise adopting a Chinese 600B checkpoint is now making a procurement decision with a compliance footnote attached.
Frequently asked questions
How much does StepFun Step 5 Preview cost?
$1.00 per million input tokens and $2.70 per million output tokens, with a 95% cache discount that takes cached input to roughly $0.05 per million.
Is Step 5 Preview open source?
Not yet. The API is proprietary and live. StepFun has committed to releasing open weights on October 15, 2026, but has not published the license, and the Hugging Face repository is currently empty.
How does Step 5 Preview score on benchmarks?
It scored 44 on the Artificial Analysis Intelligence Index against a 24 median. On StepFun’s reported figures it hits 66.4 on FrontierFinance versus GPT-6 Astra’s 55, and 33.3 on Terminal-Bench v4 versus Claude Opus 5’s 52.3.
What is the Step 5 Preview context window?
StepFun advertises 1M tokens. Separate documentation cited by OrcaRouter indicates 350,000 tokens with a 64,000-token output cap. The discrepancy is unresolved.
How fast is Step 5 Preview?
Artificial Analysis measured 99.8 output tokens per second and 2.96 seconds to first token, both better than the comparable-model medians of 70 tokens/s and 3.70 seconds.
Why is the cheap price misleading?
The model generated 160M output tokens in evaluation against a 92M median — 74% more verbose. Cost per completed task rises accordingly, even though cost per token is low.
Who owns StepFun?
StepFun is a private Chinese company founded in April 2023 and led by CEO Jiang Daxin. Reported funding figures for 2026 conflict across outlets and the company has not confirmed a valuation.
The bottom line
Step 5 Preview is the cheapest way to buy an Intelligence Index score of 44, and that is a real achievement at $1 per million input tokens. It is not a frontier model, and StepFun’s own Terminal-Bench v4 number — 33.3 against Claude Opus 5’s 52.3 — says so more clearly than any critic could.
The trade is straightforward. If your workload is token-heavy and failure-tolerant, the 80% input discount is money on the table today. If your workload is an agent that has to finish a 40-step task without supervision, the 19-point Terminal-Bench gap will cost more than the tokens ever save.
The date that matters is October 15. A permissive 600B open-weights release would reset the floor for self-hosted inference worldwide. A research-only license would make this launch a very effective press cycle for a paid API. StepFun has six weeks to decide which company it wants to be, and the verbosity numbers suggest it still has engineering to do either way.
Sources
- Artificial Analysis — Step 5 Preview: Intelligence, Performance & Price Analysis
- StepFun — Step 5 Preview launch announcement (X)
- AI Weekly — StepFun Ships Step 5 Preview API
- OrcaRouter — What StepFun’s 600B flagship really ships
- RuntimeWire — StepFun launches a 600B agent model at $1 per million input tokens
- TokenCost — GPT-5.6 Sol API pricing
Leave a Reply