StepFun Launches Step 5 Preview: 600B MoE Flagship Opens at $1 Per Million Tokens

StepFun, one of China's six leading AI labs, launched Step 5 Preview on September 20, 2026, opening API access to a 600-billion-parameter sparse Mixture-of-Experts model designed for long-horizon agentic work. The model is priced at $1.00 per million input tokens and $2.70 per million output tokens — significantly below the median pricing for comparable reasoning models, according to independent measurements from Artificial Analysis. A 95% cache discount drops repeated input to approximately $0.05 per million tokens. The open weights are not yet public; StepFun has set October 15 as the date weights become available, and the license governing commercial deployment and fine-tuning is not yet disclosed.

The practical implication for developers is simple: Step 5 Preview is callable today but not self-hostable, and October 15 is the milestone that actually determines whether this model becomes infrastructure rather than just another API.

Read more: DeepSeek V4 Flash Vision adds multimodal agent capability at text-model prices

How StepFun Designed a 600B Model That Runs on 4.5% of Its Own Parameters

Step 5 Preview is built on a sparse Mixture-of-Experts architecture, a design that has become central to large Chinese AI lab launches in 2026. Rather than activating all 600 billion parameters on every forward pass, the model routes each token through a dedicated subset of specialized parameter blocks called experts. StepFun says approximately 27 billion parameters are active per token, which represents 4.5% of the total parameter pool.

That ratio matters for two reasons. The first is inference economics: the compute required to generate each token is proportional to the active parameters, not the total. A dense 600-billion-parameter model would mobilize effectively the entire weight set on each generation step; Step 5 Preview's sparse design allows StepFun to offer a 600-billion-parameter model's representational breadth at a fraction of that per-token compute cost. The second reason is what the ratio does not mean: it does not make deployment equivalent to running a 27-billion-parameter model. The full 600-billion-parameter weight set still has to reside somewhere in the serving system. At BF16 precision, 600 billion parameters represent roughly 1.2 terabytes of weight data before additional runtime overhead — which is why distributed inference infrastructure remains necessary and why the October 15 open-weights release will be practically significant only for teams with substantial GPU clusters.

StepFun also made an architectural decision that differs from recent large MoE models: rather than widening the network, it went deeper. The company says Step 5 Preview uses a 92-layer narrow-deep Transformer layout, according to Pandaily's coverage of the launch. The stated reasoning is that deeper stacks provide longer information paths during long-context prefill — the phase when an agent reads through tool outputs, code repositories, or financial documents before producing a response. During that phase, an agent model must integrate information across hundreds of thousands of tokens, and StepFun argues that depth rather than width helps sustain multi-hop reasoning chains. This architectural claim has not been independently verified, but it is consistent with a broader research hypothesis about transformer depth and compositional reasoning.

The model supports a 1-million-token context window, accepts both text and image input, and includes extended reasoning capabilities — what Artificial Analysis classifies as chain-of-thought processing. Output speed, as measured independently by Artificial Analysis through StepFun's own API, reaches approximately 99.8 tokens per second with a time-to-first-token of 2.96 seconds, both better than the median for comparable reasoning models in the same price tier.

What the Independent Score Actually Shows — and What It Does Not

The most reliable third-party assessment of Step 5 Preview currently available is Artificial Analysis's Intelligence Index, which scores the model at 44 on its composite scale. That score places Step 5 Preview 27th among 653 tracked models on the index, which incorporates ten evaluations including AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, and others.

Artificial Analysis calculated the model cost $922.84 to evaluate across the full index — a meaningful reference point for developers estimating production costs. The platform also flags a notable verbosity characteristic: Step 5 Preview generated 160 million output tokens during the index evaluation, compared to a 92-million-token median for comparable reasoning models. A model that reasons at length spends some of its per-token price advantage on extended outputs, which means actual workload economics depend heavily on how verbose the model's reasoning chains become in production. Task-level costs still place Step 5 Preview among the better-priced models at its capability level, but developers running output-heavy applications should account for the verbosity effect explicitly.

StepFun's own benchmark table, published alongside the launch, provides additional comparison figures, though these require careful interpretation. The company reports 67.7 on DeepSWE v1.1 and 49 on StepCodeBench, putting Step 5 Preview ahead of Kimi K3 and GLM-5.3 on those measures. StepCodeBench is administered by StepFun itself, covering 553 repositories across nine task categories and 33 programming languages. The scope is credibly broad, but buyers cannot independently replicate it. On Terminal-Bench v4, which appears in the Artificial Analysis index, Step 5 Preview's company-reported score of 33.3 trails GLM-5.3 at 41.9, Claude Opus 5 at 52.3, and GPT-6 Astra at 57.9 — a gap that represents a meaningful capability difference for agentic tasks requiring extended terminal use.

Finance is where Step 5 Preview's results are most favorable relative to US competitors. On FrontierFinance, StepFun reports 66.4 for Step 5 Preview versus 55 for GPT-6 Astra and 69.7 for Claude Opus 5. On DRACO, another finance evaluation, the reported figures are 83.3 for Step 5 Preview, 76.8 for GPT-6 Astra, and 87.6 for Claude Opus 5. These are company-administered comparisons rather than a neutral common test, but the finance advantage over GPT-6 Astra is consistent enough across two separate evaluations to be worth noting for teams building financial research agents.

StepFun also published two 24-hour autonomous agent experiments in its launch materials. In the first, the company says Step 5 Preview optimized an inference kernel to 508 TFLOPS on an Nvidia H100 after roughly 22 hours of autonomous work, compared with 493 TFLOPS for Claude Opus 5. In the second, the model ran an automated post-training loop that raised a Qwen3-30B-A3B base model's AIME24 accuracy from 53.3% to 60%, matching Claude Opus 5 while using fewer annotator tokens. These demonstrations are compelling as directional evidence for sustained agentic execution, but the experimental setups were defined and run by StepFun, with limited methodological detail about the grading process. A separate claim — that roughly 70% of internal and external experts judged the model capable of autonomously completing moderately complex coding tasks — is similarly limited in verifiability because the participant count, task set, and evaluation rubric were not disclosed.

Step 5 Preview in the Context of the Chinese Frontier Model Race

StepFun is one of China's six major AI labs, sometimes called the "Six Tigers." Jiang Daxin, who spent 16 years at Microsoft working on Bing, Cortana, and Azure Cognitive Services before earning a computer science PhD from the University at Buffalo, founded StepFun in April 2023 shortly after ChatGPT catalyzed the generative AI wave, according to a 2025 profile in The Wire China. The company has received strong institutional backing, with a Series B+ round exceeding $718 million announced in January 2026, backed by Tencent, Qiming Venture Partners, and Shanghai state-owned capital funds, as reported by TechNode.

Step 5 Preview lands in a market where the benchmark competition among Chinese frontier labs is intensifying. Moonshot AI released Kimi K3 in July 2026 — a 2.8-trillion-parameter sparse MoE model with 104 billion active parameters and a 1-million-token context window that achieved an Artificial Analysis Intelligence Index score of 44, matching Step 5 Preview on that measure. Alibaba has previewed Qwen3.8-Max with 2.4 trillion total parameters. Z.ai's GLM-5.3 outperforms Step 5 Preview on Terminal-Bench v4 despite a smaller headline parameter count.

The most important competitive fact is one of product strategy rather than benchmark score. Kimi K3 released its full open weights by July 27, 2026, under a permissive distribution. Step 5 Preview has not yet done this. For the approximately three and a half weeks until October 15, Step 5 Preview is a 27th-ranked proprietary API model, and the community's evaluation of it is limited to what can be inferred from the hosted API. Once weights land, the assessment changes materially: independent researchers can measure memory requirements, evaluate quantization behavior, attempt fine-tuning, and apply whatever licensing terms apply to determine whether commercial derivative works are permitted.

StepFun is running a deliberate dual-track strategy. On the open side, Step 3.5 Flash — a 196-billion-parameter sparse MoE model with 11 billion active parameters and a 256,000-token context window — was released in February 2026 under the Apache 2.0 license, making it freely deployable and modifiable for commercial use, as confirmed by Artificial Analysis's model data. Step 5 Preview sits above that, as a proprietary closed flagship that StepFun serves through a metered API. The structure mirrors how other Chinese labs are managing the tension between community adoption and commercial revenue, with open smaller models building developer trust and closed flagship models capturing enterprise spending.

Pricing, Caching, and the Real Economics of Agent Workloads

The $1 input and $2.70 output pricing is independently verified by Artificial Analysis, which places both figures well below the platform's medians of $1.88 input and $10.00 output for comparable reasoning models. The effective input price with caching applied is approximately $0.05 per million tokens — a figure relevant for any agent architecture where large portions of the context repeat across turns.

Agent workloads are often highly cache-amenable. A coding agent working from a fixed codebase sends the same repository headers, system instructions, and project context on every turn. A financial research agent repeatedly references the same set of filings or policy documents. If 90% of a million-token context is stable system-prompt material, the cached input price at $0.05 per million makes cost-per-task estimates far more favorable than the listed $1 input price implies.

Read more: Alibaba launches Qwen3.8-Omni-Flash as Chinese AI agent pricing falls below $0.15

That said, the verbosity characteristic works in the opposite direction. Reasoning models that chain-of-thought extensively before producing a final answer generate large volumes of output tokens. Step 5 Preview's 160-million output token usage in the Artificial Analysis index evaluation — versus a 92-million-token median for comparable models — suggests the model reasons at roughly 1.7 times the typical length. At $2.70 per million output tokens, extended reasoning on each task erodes the per-token input advantage. Developers building high-throughput pipelines need to measure the verbosity effect in their specific task distributions before projecting cost savings.

The model also exposes an OpenAI-compatible API endpoint, lowering the integration work for teams that already use that client format. Rate limits, geographic access restrictions, and enterprise-tier pricing have not yet been published.

What October 15 Actually Changes

The API launch on September 20 makes Step 5 Preview callable and independently benchmarkable through hosted infrastructure. It does not make the model deployable. The October 15 release changes that — but only if several conditions hold that cannot currently be evaluated.

First, the license terms. "Open weights" is a broad description that can cover anything from an Apache 2.0 license permitting unrestricted commercial use to a custom agreement that prohibits derivative models, restricts deployment geography, or bans commercial fine-tuning. Until the actual license is published on October 15, teams evaluating Step 5 Preview for enterprise deployment cannot fully assess whether self-hosting is legally viable for their use case.

Second, the deployment requirements. As noted by Orca Router's analysis of the announcement, the Hugging Face repository stepfun-ai/Step-5-Preview-BF16 already exists but contains only a .gitattributes file — no weights, no model card, no configuration. When weights arrive, teams will need to assess the quantization options available, the supported inference stack, multi-node requirements, and the practical throughput achievable on hardware they actually operate. The 600-billion-parameter weight set, even sparsely activated, requires substantial memory bandwidth and high-speed interconnect for efficient multi-node serving. Whether StepFun will release optimized quantized variants alongside BF16 weights is unknown.

Third, fine-tuning access. Sparse MoE models present technical complications for standard fine-tuning procedures: parameter-efficient methods like LoRA need adaptation for expert-gated architectures, and full fine-tuning on a 600-billion-parameter model is beyond the resources of most organizations regardless of licensing. Whether StepFun will release fine-tuning guidance, modified architectures suitable for adaptation, or distillation-friendly checkpoints is an open question.

The step that matters most for the research and development community is therefore not the API launch — it is what happens in three and a half weeks. The API makes Step 5 Preview a new price point in the agentic model market. The open-weights release will determine whether it becomes a foundation others can build on.