StepFun's Step 5 Preview API ships; open weights promised October 15

StepFun shipped the Step 5 Preview API on September 20 and says the model's weights go open on October 15. Until then it is a closed preview: a 600B-parameter sparse mixture-of-experts with 27B active per token, priced at $2.70 per million output tokens, with every benchmark number on the page StepFun's own.

Key facts

• 600B total parameters, 27B active per token (4.5%), 1M-token input context, 64k output tokens, per StepFun's model documentation.
• List price $1.00 per million input tokens on a cache miss, $0.05 on a cache hit, $2.70 per million output tokens, reasoning tokens included — all StepFun's own pricing.
• Open weights promised for October 15, 2026; no license named, and no Step 5 repository existed under the stepfun-ai Hugging Face organization a day after launch.
• StepFun reports 67.7% on DeepSWE v1.1 and 49.0% on its own StepCodeBench, running at High effort against rivals' Max modes.
• StepFun cites a score of 44 on the Artificial Analysis Intelligence Index — a claim to be confirmed on Artificial Analysis's own page, not by the vendor citing it.

What happened

StepFun announced Step 5 Preview at 03:15 UTC on September 20 on stepfun.com, went live in its own products and API under the model id step-5-preview, and dated the open-weights release October 15 — 25 days out. The announcement's title is "Advancing the Pareto Frontier," and the claim it makes is economic rather than top-line: comparable intelligence to the current open-weight flagships at a fraction of the per-task cost.

The architecture is documented as a sparse mixture-of-experts with 600B total parameters and 27B active per token — the smallest active count in StepFun's own comparison table, where Moonshot's Kimi K3 activates 104B of 2.8T parameters. Context is 1M tokens input with 64k output; inputs cover text, up to 60 images per request, and video (MP4, QuickTime, Matroska). Reasoning effort is set per request at low, medium, or high, and the endpoint is OpenAI-compatible.

The pricing is the part that matters to builders. StepFun lists $1.00 per million input tokens on a cache miss and $2.70 per million output tokens, with reasoning counted as output. Against the models StepFun benchmarks itself against, that output price is a fraction of Kimi K3's $15.00 list price on Moonshot's own API, and several times more than flash-class open-weight endpoints like GLM-5.3-Flash at $0.50. The cost claim is checkable on list prices; whether the intelligence is comparable is not, because the benchmark table is StepFun's.

That table has three caveats StepFun states itself: Step 5 Preview ran at High effort while GPT-6 Astra, Claude Opus 5, Kimi K3, and GLM-5.3 ran at Max; six rows are StepFun's own benchmarks, including StepCodeBench; and one HLE-with-tools row mixes a text-only subset against full-dataset runs that StepFun itself calls not directly comparable. On the rows that stand, Step 5 Preview runs level with Kimi K3 and GLM 5.3 on most measures — 67.7% on DeepSWE v1.1 against K3's 67.5% and GLM 5.3's 66.9% — and six to seven points behind GPT-6 Astra and Claude Opus 5 on the hardest coding rows.

The most interesting evidence is not the table. StepFun ran three long-horizon experiments: 24 hours optimizing an MLA GPU kernel on an H100 to 508 TFLOPS against a reported 493 for Claude Opus 5, 24 hours of automated post-training lifting a Qwen3-30B-A3B base from 53.3% to 60% on AIME24, and 3,000-plus turns of Pokémon Red. A vendor-run 3% kernel edge is a data point, not a verdict — but the shape of the pitch is clear: a model priced so that a day of autonomous iteration is affordable.

Why it matters

Nothing here is downloadable, and that is the whole story for anyone deciding where to run long-horizon agent workloads. A 600B-total/27B-active model with a 1M-token context is exactly the architecture profile that changes the rent-versus-host math for agentic serving: the active-parameter footprint is what you pay to keep resident, not the 600B. Today the only way to touch the model is StepFun's API at StepFun's prices. If the October 15 release lands as promised, third parties can host it, quantize it, and price it — the same path Kimi K3 took after its July 27 open-weights release, which put a 2.8T-parameter model in third-party serving stacks within weeks.

The claim to verify is narrower than it looks. StepFun says "open weights" on October 15; it has not named a license, linked a technical report, or published a Hugging Face repository. A license alone can flip the value of a weights drop in either direction: Qwen's Qwen-Image-2.1 shipped open weights on September 20 under a research-only license that blocks commercial use without a separate agreement — our PULSE coverage walks through that text. Open weights are a delivery event; what ships alongside them decides whether production users can act on it. That distinction between weights and rights is the recurring lesson of our open-weight-versus-open-source explainer.

There is also a market-structure angle. The Institute of Foundation Models' K2 Horizon family and Moonshot's Kimi K3 both made open-weight flagships a normal supply event this quarter (our K2 Horizon coverage). StepFun is doing the opposite sequence — closed preview first, weights later — which lets it collect API revenue and tune serving before anyone else can host the same checkpoint. Whether the October 15 drop includes serving-side documentation good enough for a third party to match its throughput claims is the practical test of how open the release really is.

Background

StepFun is a Shanghai-based frontier lab whose prior open releases — Step 3.5 Flash and Step 3.7 Flash — are listed on OpenRouter, making it one of the few Chinese labs whose models are directly reachable through Western aggregators. The pricing posture in this launch follows the pattern that has defined Chinese frontier-lab pricing through 2026: Kimi K3 set the flagship anchor at $15.00 per million output tokens on Moonshot's own API, GLM-5.3-Flash and DeepSeek V4.1 Flash undercut from below at $0.50 and $0.60, and Step 5 Preview slots a flagship price of $2.70 between them. These are each vendor's own list prices, not an independent survey.

Two things would verify the launch beyond StepFun's word. First, the October 15 repository — an actual stepfun-ai Hugging Face organization entry with a named license, the same standard we applied to Kimi K3's open-weights release. Second, independent numbers: Artificial Analysis's own index page for the model, and reproduction of the DeepSWE row by anyone not employed by StepFun. The company cites a score of 44 on the Artificial Analysis Intelligence Index, but at publish time the confirmation has to come from Artificial Analysis, not from the launch page quoting it.

It is worth holding both halves of this launch in view at once. The pricing pressure is real and checkable today — $2.70 output against $15.00 for the closest open-weight flagship is the kind of gap that forces repricing across the market if the capability claim holds. The openness promise is not checkable until October 15. Treating the first as fact and the second as a date on a slide is exactly the discipline this beat requires.

What's next

October 15 is the date that converts this from announcement to release. Between now and then, watch for three things: a license named on the announcement page or model card, a stepfun-ai Hugging Face repository, and a third-party row on OpenRouter — the place where StepFun's list price meets actual market pricing. If any of the three slips, the open-weights promise moves from delivery to another preview.