StepFun's Step 5 Preview: a 600B-parameter agent model at $1 per million tokens, open weights land October 15

Shanghai has produced another giant model — but this time the headline is the price tag. On September 20, the Chinese lab StepFun opened API access to Step 5 Preview, a 600-billion-parameter sparse mixture-of-experts model it describes as a flagship built for agentic work: software engineering, professional knowledge tasks, and finance. The pitch is not that it is the smartest model alive. It is that it delivers near-frontier results at roughly a dollar per million input tokens — about a quarter of what several closed frontier labs charge for comparable workloads.

The framing is deliberate. StepFun's launch tagline is "Advancing the Pareto Frontier": more capability per dollar, not more parameters per headline. Where the summer's open-weight shockwaves were scale stories — Kimi K3's 2.8 trillion parameters, Xiaomi's MiMo-V2.6 taking the top of the open-weights leaderboard — Step 5 Preview is an economics story, and one aimed squarely at builders who want to run agents overnight without watching the meter spin.

What StepFun actually shipped#

Step 5 Preview is a sparse mixture-of-experts model with 600 billion total parameters and only 27 billion active per token — about 4.5% of the weights engaged for any given output. That sparsity is the lever behind the pricing: cost tracks what computes, not what sits in memory. The model carries a 1-million-token context window, takes text, image, and — per the official documentation — video as input, and supports streaming, tool calling, JSON mode and schema, and prompt caching. StepFun also exposes three reasoning-effort levels (low, medium, high) and says one agent run coordinated 950 web fetches in a single action — a company-reported figure worth treating as marketing until reproduced, but directionally consistent with the model's agentic positioning.

For now it is API-only, live on StepFun's platform under the model ID step-5-preview. Full open weights are scheduled for October 15, with the license still unannounced — an important caveat for anyone planning to self-host. As a rough guide to the hardware question, 600 billion parameters is about 1.2 terabytes in BF16 before any KV cache, so self-hosting will be a multi-GPU-server affair regardless of the license.

The benchmarks — read them twice#

The independent numbers come from Artificial Analysis, which scores Step 5 Preview at 44 on its Intelligence Index, ranks it 27th of 653 models, and calculates a cost of $0.71 per Intelligence Index task. One independent outlet notes the model ties Kimi K3 (max) on that index at roughly 65% lower cost per task — the cleanest single-line summary of StepFun's pitch. The same evaluation records an output speed of 99.8 tokens per second and a 2.96-second time to first token.

The company's own numbers need the usual discount. StepFun reports 67.7% on DeepSWE v1.1, 49.0% on its internal StepCodeBench, and 66.4 on FrontierFinance, and secondary reporting describes it as trailing only GPT-6 Astra or Claude Opus 5 among evaluated systems — ahead of other open-weight entrants. StepFun presents StepCodeBench as a 553-repository, nine-category, 33-language suite; that it is a StepFun-designed benchmark evaluating a StepFun model should temper enthusiasm until outside labs reproduce the runs. Notably, no SWE-bench Verified score exists yet, so the model enters at least one independent coding leaderboard as unverified.

The price is the point#

Step 5 Preview's API pricing is $1.00 per million input tokens, $2.70 per million output tokens, with cached input discounted 95% to about $0.05 per million. StepFun says the per-task inference cost runs roughly one-eighth of Claude Opus 5's — a company claim, but directionally backed by the independent cost-per-task numbers above.

That economics matters more than any leaderboard position. The open-weight frontier in 2026 is no longer really about who tops a static index; it is about whether cheap, long-context inference is good enough to run real agent loops — overnight coding agents, research pipelines, back-office document processing — at costs that make deployment an obvious call instead of a budget fight. A 1-million-token context window is enough to hold a large repository, a research pack, and the intermediate results of a long tool-using session in one sitting. If the agentic behavior holds up, the model that is "good enough" at a dollar a million tokens changes more buying decisions than the model that is 2% smarter at eight times the price.

China's open-weight wave#

StepFun — 阶跃星辰 — was founded in Shanghai in April 2023 and is counted among China's "Six AI Tigers," the independent large-model startups alongside Z.ai, Moonshot AI, MiniMax, Baichuan, and 01.AI. Its earlier open releases, Step 3.5 Flash and Step 3.7 Flash, shipped under Apache 2.0, which is why the unannounced license for Step 5's weights is worth watching: the lab has precedent for openness, but precedent is not a guarantee.

The release lands in a year when Chinese labs have repeatedly collapsed the assumed gap with the American frontier — DeepSeek's reasoning shock, Kimi K3's top-3 open-weight rank, Xiaomi's trillion-parameter open-sourcing. Each wave has shortened the distance between "open model from China" and "frontier capability." Step 5 Preview continues the pattern on a different axis: not more scale, but more capability per dollar.

What to watch#

Three things, in order. First: October 15. Do the weights actually ship, and under what license? Until they do, everything about Step 5 Preview is an API business, and the open-weights framing is a promise. Second: independent verification of the agentic claims. Company benchmarks are press releases with error bars; the SWE-bench Verified absence is the tell. Third: the pricing response. If a Chinese lab can sell near-frontier agent inference at $1 per million input tokens with real traction, the closed labs will have to answer — and the most likely answers are cheaper models or better agents, both of which benefit builders either way.