Step 5 Preview's 1M-Token Window Just Left StepFun's Own API

Step 5 Preview stopped being a single-vendor model on October 16, 2026. StepFun's own account confirmed that day that its 1M-token flagship is now live on third-party inference hosts, which is the first time the model has been reachable by a developer who does not want a StepFun account, a StepFun balance and a StepFun rate card in a second currency. That is the entire news value of the announcement and it is worth stating plainly, because the announcement itself carried no new specification, no price change and no benchmark. It carried a distribution change. StepFun released Step 5 Preview on September 20, 2026; what happened on October 16 is that the model started showing up somewhere other than home. That gap between what a scoreboard measures and what a deployment meets is the reason the record matters: the documented AI failures and agent incidents around models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Distribution changes are easy to underrate and they are the ones that move procurement. A model available from exactly one vendor is a single point of failure, a single set of rate limits and a single set of regional endpoints, and the honest answer to "what happens if that vendor has a bad afternoon" is "your product has a bad afternoon." A model available from several independent hosts is a routing problem, and routing problems have solutions. The catch is that not everything about Step 5 Preview travels with the weights to a new host, and the thing that does not travel is the number most likely to decide your bill.

What StepFun actually said, and what it left out

The announcement is one sentence long in substance: the model is live on third-party inference platforms, and the 1M-token context window is the feature being emphasised. It does not name a price on those platforms, does not commit to a context window being served at full length, and does not say anything about whether the prompt-cache discount follows the model across the boundary. Those three omissions are the three questions a team should ask before moving a pipeline, and they are not pedantry — each of them has a plausible wrong answer.

The model being distributed is the same one StepFun documented in September, and the documented figures have not changed. From StepFun's own model page:

• Model ID — step-5-preview, with "Preview" part of the identifier rather than a suffix this article added

• Context window and maximum input — 1M tokens each, per StepFun's specification table

• Maximum output — 64k tokens, with max_tokens documented as defaulting to unlimited so the model chooses its own stopping point unless you set one

• Input — text, images and video; images as URL or Base64 in JPG, PNG, WebP or static GIF up to 60 per request, video as MP4, QuickTime or Matroska

• Output — text only, with streaming, tool calling, JSON Schema structured output and reasoning effort at low, medium or high

• Architecture — a sparse mixture-of-experts model at 600 billion total parameters and 27 billion active per token, per StepFun's launch material and the model card mirrors that carry it; StepFun's own documentation page does not print a parameter count

• Price on StepFun's own platform — $1.00 per million input tokens, $0.10 per million cache-hit input tokens and $2.70 per million output tokens

None of that is new. It is the baseline the October announcement is distributing, and it is worth restating because a third-party host is free to serve a different quantisation, a shorter effective window or a different cache arrangement, and none of those choices appear in a launch post.

Why the 1M window is the part being sold

The 1M-token context is the reason this model has a distribution story at all. StepFun built Step 5 Preview around long-horizon agentic work, and the window is what makes a repository-scale or document-set-scale prompt possible in a single call rather than through a retrieval layer you have to build and maintain. On the independent board, Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index and computes a cost of roughly $0.71 per completed index task — a figure that folds verbosity into the price and is therefore less flattering than the rate card. Output speed readings on the same page have moved between captures, from 88.1 to 99.8 output tokens per second, which is itself a reminder that hosted throughput is a property of the host as much as of the model.

That last point is the one the announcement makes relevant. A 600-billion-parameter sparse model with 27 billion active parameters is expensive to serve at a million tokens of context, because the KV cache for a long prompt is a real memory cost per concurrent request. Whoever serves it has made a decision about how much of that window to honour, how many requests to run at once, and whether to quantise the attention stack to fit more of them on a card. StepFun's own FP8 checkpoint is 650 GB on disk and its BF16 checkpoint is 1,215 GB; a host running either is running a different machine from the one StepFun runs, and the throughput a vendor quotes for its own cluster is not a promise about somebody else's.

The number that does not travel: the cache discount

Step 5 Preview's rate card has a 95% discount on cached input, which takes a re-read token from $1.00 per million to $0.10. For a single-turn request this is irrelevant. For the workload the model is actually sold for — an agent loop that resends the same system prompt, tool schema and repository map on every turn — it is the dominant term in the bill, and it is the reason Step 5 Preview's effective rate on a long session is nowhere near its sticker.

Prompt caching is an implementation detail of a serving stack, not a property of weights. Whether a given host offers it, how it identifies a cache hit, how long a prefix stays warm and what it charges for the read are all host decisions. StepFun documents caching on StepFun's platform. An announcement that a model is live somewhere else tells you nothing about it, and this is the specific line to read on a new host's pricing page before assuming the economics you modelled on StepFun's card still hold. A host that charges full input price on every turn can be several times more expensive than StepFun for the same workload while looking cheaper per million tokens on the page.

StepFun's own launch-week distribution has been unusually broad for a Chinese lab's flagship, and it has arrived in stages rather than at once: the model appeared in third-party coding tools on a free trial for a week in early October, and the third-party API hosting announced on October 16 is a later and more general step. Read the two together and the pattern is a lab that wants the model in front of developers wherever they already are, rather than driving them to its own console. That is a coherent strategy and it is also a strategy with an obvious cost: every additional host is another place where the model's envelope is somebody else's implementation.

Three questions worth asking a new host

The useful response to this announcement is not to switch, it is to check. These are the three items that decide whether a new endpoint is the same product.

• Quantisation — ask which checkpoint is being served. StepFun publishes FP8, BF16 and, since October 14, an NVFP4 build at 394 GB, and those three are not interchangeable in quality or in memory footprint. A host running NVFP4 on Blackwell is running a four-bit expert stack with attention and the routing gates held back; whether that costs you accuracy on your task is a question only your own eval answers.

• Effective context — ask whether 1M tokens is served to every caller or capped per tier. Serving a million-token prompt has a concurrency cost that shows up as a rate limit before it shows up as a price, and a host that advertises the window while throttling the concurrency is advertising the model's document rather than the service.

• Cache semantics — ask what a repeated prefix costs. This is the line that decides long-session economics, and it is the line most likely to be omitted.

A run against the same prompt on both endpoints, measured in dollars per completed task rather than dollars per million tokens, answers all three at once and takes an afternoon.

Where this leaves the model

Step 5 Preview is a 600-billion-parameter sparse mixture-of-experts model with a million-token window, text, image and video input, a 64k output ceiling, a $1.00 / $2.70 rate card and a 95% cache discount on StepFun's own platform, scored at 44 on the one independent index that covers it. Released September 20, 2026. As of October 16 it is also reachable, in some form and at some price, from at least one host that is not StepFun.

It is worth saying plainly that Step 5 Preview is not in OrcaRouter's live catalogue, so the place to call it remains the vendor's own API or a third-party host — and if you are evaluating it, that gap is the argument for keeping the rest of your stack on a single endpoint. One API covering 200-plus models at provider list price with no markup is what lets a two-week experiment on a new model stay an experiment: your fallback is a routing rule rather than a second integration, automatic failover means an unfamiliar endpoint cannot take your production path down with it, and because the pass-through is the provider's own price, a vendor's rate change reaches your bill the day it lands instead of on the next contract cycle. The model you are testing is the risk; the plumbing around it should not be a second one.

The next thing to watch is not another host. It is the first host that publishes its own throughput, its cache terms and its quantisation on the same page, so the comparison stops being a guess about what the 1M-token window costs when somebody other than StepFun is paying for the GPUs.