Step 5 Non-Preview: What StepFun's Final Flagship Has to Beat

The entire public record of a model people are calling Step 5 Non-Preview, as of the morning of 6 November 2026, is one sentence posted to X at 15:47 UTC the previous day: "I really hope Step 5 Non-Preview is strong." It has twenty likes. Nothing else about this model exists that can be checked — no repository, no model ID in StepFun's documentation, no entry on any leaderboard, no date, and no statement from StepFun itself. Step 5 Preview, the build that does exist, shipped its API on 20 September 2026 and its weights on 15 October; the unreleased successor has, so far, a name the community invented and a hope attached to it. This page is what we can verify about both, and what a final Step 5 would have to clear if it ever lands.

The one post, read carefully

The post is from @teortaxesTex, an English-language account whose profile names DeepSeek, a Chinese lab, as its subject — not a StepFun account, and not a StepFun employee. It quotes a post published the same morning by @liulicheng10 — "my passion for my work stems from solving open-ended, challenging, and scientific problems" — and adds the hope about Step 5 Non-Preview. That is the whole of it. We could not establish from public sources that the quoted account belongs to a StepFun employee, and the quoted post says nothing about a model, a date or a feature. Treat any reading of it as a staff hint as unverified. Two things are nonetheless worth noting about the signal. First, the phrase "Non-Preview" is being used as if it names a specific, expected thing rather than a wish — which is how a community talks about a successor it believes is coming, not about a hypothetical. Second, "twenty likes" is the correct measure of its evidentiary weight. It tells you what one well-read observer expects. It does not tell you what StepFun is doing.

This is the weakest kind of early signal this blog covers, and that is the honest frame. It is not a framework pull request that names a new architecture, not an arena codename that leaked ahead of a launch, and not a staff post that can be tied to a person. It is a sentence about a model that does not exist yet.

What exists, and what does not, as of 6 November 2026

Five independent checks, each of which would normally be the first place a new StepFun model shows up. All five are negative for a non-preview Step 5, and all five are positive for Step 5 Preview.

• StepFun's own public model list. The documentation's Model Overview page, read today, lists Step 5 Preview as the flagship with "API model ID: step-5-preview" and no second Step 5 entry. The publisher's full documentation export contains that identifier more than two hundred times and a bare step-5 not once.

• StepFun's Hugging Face organisation. Queried through the Hugging Face API today, stepfun-ai holds three Step 5 repositories — Step-5-Preview (FP8), Step-5-Preview-BF16 and Step-5-Preview-NVFP4 — created between 20 September and 14 October 2026. There is no repository named Step-5, and nothing created in November.

• StepFun's GitHub organisation. The stepfun-ai/Step-5-Preview repository was created on 12 October 2026 and contains a README, two licence files and a contributors list — documentation only, no weights. No successor repository exists.

• Artificial Analysis. The evaluator tracks five StepFun models: Step 3.5 Flash 0202, Step 3.5 Flash 2603 and Step 3.7 Flash (all flagged deprecated), Step3 VL 10B, and Step 5 Preview — which is flagged as not deprecated and carries a release date of 18 September 2026. There is no non-preview Step 5 in its model list.

• Arena listings. The only Step 5 entries that have surfaced in arena tracking this autumn carry the preview suffix, including a step-5-preview-agent entry caught on 2 October 2026. A bare step-5 has not appeared.

One more absence matters: StepFun's Model Lifecycle and Deprecations page, the document that would carry a retirement notice for step-5-preview and a replacement mapping, does not mention Step 5 Preview at all. Its most recent entries are the 8 July 2026 retirement of nine older identifiers and the 10 October 2026 retirement of the image model. So there is no published expiry date on the preview either — which means the current state is not "waiting for the preview to be replaced", it is "no lifecycle event has been announced in either direction".

Why "non-preview" is a different model, not a rename

StepFun has not said what a final Step 5 would be, so the only useful thing to establish is what a change of that kind has meant at this vendor before. Two documented patterns apply.

The first is that the model ID is the contract. step-5-preview is not a marketing label bolted onto a product name; it is the exact string a caller sends in the model field, and it appears as the identifier in the documentation table, the pricing table and the Hugging Face card. Whatever a final build is called, step-5 and step-5-preview cannot be the same string unless StepFun changes the identifier on the preview itself — in which case existing integrations break on the day of the change, not on the day of a retirement.

The second is how StepFun ends identifiers. The July 2026 retirement record lists nine models retired on one date, each with a recommended replacement and an instruction to change the model field — step-3, for example, was replaced by step-3.7-flash. Calls to retired IDs do not return a warning; they stop returning valid responses. So the two failure modes a reader should plan for are an ID that stops existing and an ID that quietly starts behaving differently, and StepFun's own documentation addresses only the first.

Neither pattern tells us whether a final Step 5 keeps the preview identifier, takes a new one, or arrives as a differently sized model under a new name. Nobody outside StepFun knows, and the post that started this page does not say.

The bar: what a final build has to clear

If a final Step 5 ships, it will be judged against the preview, so the preview's numbers are the benchmark — read today, at their current values, because several numbers in circulation for this model are stale.

• Size and shape: 600 billion total parameters with 27 billion active per token, a sparse mixture-of-experts design, per StepFun's model card. The card also describes a separate vision encoder, and its serving commands enable multi-layer EAGLE speculative decoding.

• Context: up to 1 million tokens, with text, image and video input and text-only output.

• Maximum output: this is where the first-party documents disagree. StepFun's platform documentation gives a maximum output of 64k tokens. The Hugging Face model card's specification table gives "Up to 1M tokens". Both are StepFun documents. A deployment note in that same card says the 1M-token limit covers input and output combined, which cannot leave a full 1M tokens for output alone — so the card contradicts itself, and the platform documentation is the more conservative of the two. If you are sizing an output budget, plan against 64k until StepFun resolves it.

• Score: an Artificial Analysis Intelligence Index of 29 on v5.0.1, at $2.35 per index task. The 44 that October coverage of this model quotes is a reading from the pre-v5 rubric, which Artificial Analysis replaced with v5 in late October 2026; the two are not comparable and should never be subtracted from one another. Anyone repeating 44 today is quoting a rubric that no longer exists.

• Price: $1.00 per million input tokens on a cache miss, $0.10 per million on a cache hit, and $2.70 per million output tokens, per StepFun's English pricing page. The Chinese platform lists the same model at 7, 0.70 and 20 yuan, and raised the cache-hit rate from 0.35 to 0.70 yuan at 00:30 Beijing time on 14 October 2026, grandfathering callers who had already used the pay-as-you-go API.

• Licence: the StepFun Model License 1.0, described by StepFun as Apache-2.0-style with added conditions. Two of those conditions are commercial: attribution ("Powered by StepFun AI") becomes mandatory once a product reaches more than 100 million active users in a month or more than $20 million in revenue, and a separate commercial agreement is required for hardware makers shipping the model as a device feature, or for anyone running a model-as-a-service business above $20 million in combined revenue.

That is the bar. A final Step 5 that scores below 29 on the current index, or costs more per task than $2.35, would be a harder sell than the preview it replaces — which is presumably the substance behind the hope in the original post.

What would count as a real signal

The gap between a post like this one and a launch is five observable events, in rough order of how early they tend to appear. Each is checkable in under a minute, and none has happened.

• A new entry in StepFun's Model Overview page, or a step-5 identifier appearing in the pricing table next to step-5-preview.

• A repository under stepfun-ai on Hugging Face whose name drops the preview suffix, or a Step 5 repository created after 14 October 2026.

• A model page in Artificial Analysis or a fresh arena entry without the preview suffix. The September build surfaced on the evaluator before StepFun's own announcement, so this channel has form.

• An entry on the Model Lifecycle and Deprecations page — a retirement date for step-5-preview, or a replacement mapping that names a successor.

• A dated vendor post from StepFun's own accounts or its technical blog, of the kind it published for the preview's weights on 15 October.

Until at least one of those lands, "Step 5 Non-Preview" is a phrase describing an expectation, and a piece like this one is the correct amount of coverage: the verified state of the model, not a launch story.

If you are building on Step 5 Preview today

Three practical notes, all of which follow from what StepFun has and has not published.

Pin the identifier explicitly rather than letting a default resolve it, and keep the string in configuration rather than in code, because the likeliest disruption from a final Step 5 is a change to that string or to what it points at. Watch the lifecycle page rather than the marketing pages: it is the only StepFun document that has ever carried a date at which calls stop working, and it currently says nothing about this model. And keep a second model reachable behind the same OpenAI-compatible client, so that a change on StepFun's side is a routing decision rather than an outage.

That last point is the honest place to mention what we do. Step 5 Preview is not on OrcaRouter's live model list as of 6 November 2026 — the catalogue carries no StepFun model at all, and we are not going to imply otherwise. What the router is for in a situation like this one is the other half of the hedge: the model you fall back to can sit behind the same endpoint and the same key as the models you already run, with provider list price passed through at 0% markup rather than repriced on the way. If a final Step 5 later appears on a provider we route, moving to it is an identifier change and a routing rule, not a second contract and a new SDK. Today it is not there, and the fallback side is the part we can actually do something about.

What would make this page wrong

A repository appearing under stepfun-ai with a name that drops "Preview". A model ID showing up in StepFun's pricing table beside the current one. A ranked entry on the evaluator without the suffix. Any of those turns this from a what-we-know page into a launch story, and the numbers above become the comparison rather than the subject. Until then, the honest summary is the one-line version: Step 5 Preview is real, dated and priced; a final Step 5 is a hope with twenty likes, and the gap between those two things is wider than the name suggests.