Shanghai-based StepFun has released Step 5 Preview, a 600-billion-parameter sparse mixture-of-experts model with a one-million-token context window, priced at $1.00 per million input tokens — a fraction of what Western frontier labs charge for comparable long-horizon agentic work. Announced on 20 September 2026 and already live in StepFun's products and API, the model is positioned squarely at the workloads enterprises actually struggle to afford: multi-hour coding sessions, software engineering, financial analysis and professional knowledge work.
The Architecture: Narrow and Deep
Step 5 Preview activates roughly 27 billion parameters per token out of about 600 billion total — approximately 4.5% of weights on any given forward pass. That ratio is aggressive even by current MoE standards, and it is what makes the pricing possible.
The more unusual decision is structural. Rather than widening the network, StepFun built a 92-layer narrow-and-deep Transformer. The company's argument is that deeper stacks create longer information paths, which support implicit multi-hop reasoning during long prefill — precisely the regime a million-token context creates. It is a bet that depth, not width, is the right lever for long-horizon agentic reasoning.
The model accepts text and image inputs, and reasoning effort is selectable per request at low, medium or high against the same model ID, letting developers dial cost against difficulty without switching endpoints.
Notably, StepFun skipped the Step 4.x line entirely, jumping from Step-3.7-Flash straight to Step 5.
Pricing Is the Headline
On StepFun's API, the model costs:
- $1.00 per million input tokens on a cache miss
- $0.05 per million input tokens on a cache hit
- $2.70 per million output tokens, with reasoning tokens billed as output
The company positions this as under a fifth of comparable OpenAI pricing. The 20x cache-hit discount is the number that matters most for agent builders, since long-running agents replay enormous shared prefixes — system prompts, tool schemas, retrieved documents — on every turn.
Where It Sits on the Benchmarks
Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, placing it 24th among roughly 200 tracked models against a median of 24. That is a strong result for the price band, without being a frontier claim.
StepFun's own figures report 67.7 on DeepSWE v1.1, 49 on StepCodeBench and 80.5 on ProgramBench, ahead of Kimi K3 and GLM-5.3 in the company's comparison table. On the same tests, GPT-6 Astra and Claude Opus 5 remained ahead. In other words: not the best model available, but arguably the best available at this cost per unit of capability.
One dating discrepancy is worth noting for readers tracking release chronology — Artificial Analysis lists the model as of 18 September, two days before StepFun's own announcement.
The Open-Weight Promise, and Its Asterisk
StepFun says weights will be released on 15 October 2026. That commitment is the strategic core of the announcement: it would make Step 5 one of the largest openly available models with a million-token context.
The caveat is that, as of publication, it remains a promise. The Hugging Face repository `stepfun-ai/Step-5-Preview-BF16` existed within hours of the announcement but contains a single `.gitattributes` file — no weights, no licence, no model card, no configuration. No licence has been named in the announcement and no technical report has been linked. Until October, this is a closed preview carrying an open-weight pledge.
Teams planning to self-host should also do the arithmetic now. At 600 billion parameters in BF16, the weights alone require roughly 1.2 TB before any KV cache, putting inference firmly in multi-GPU server territory. A million-token context makes the KV cache burden substantially worse.
Why It Matters
Step 5 Preview is the clearest expression yet of a strategy several Chinese labs are converging on: concede the absolute capability crown, then compete relentlessly on cost per useful token.
That strategy has three consequences for the wider market.
It compresses the price floor for agentic work. Long-horizon agents are economically viable only when the marginal cost of a reasoning step is small. Pricing at $1 per million input tokens, with a five-cent cache hit, moves workloads that were previously uneconomic into range.
It pairs with a dual-model playbook. StepFun released Step 3.5 Flash in February as an open-weight MoE with 196B total and 11B active parameters at 256K context, aimed at local deployment. Step 5 is larger, currently proprietary and metered as a service. The lab now has an open model developers can run themselves alongside a closed flagship it can monetise — the same structure several Western labs have adopted in reverse.
It raises the stakes for October. If the weights ship on schedule with a permissive licence, Step 5 becomes a serious option for regulated enterprises that cannot send long documents to a third-party API. If they slip, the announcement will be read as marketing. The 15 October date is now the thing to watch.
