Step 5 Preview Is Free in Kilo Code for One Week — and That Week Is Documented Nowhere

Step 5 Preview is free inside Kilo Code for one week. That sentence is the entire announcement, and it is the only place the length is written down. The model itself is not new — StepFun announced it on 20 September 2026, twenty days before this promotion, and it already costs $1.00 per million input tokens and $2.70 per million output tokens on the vendor's own platform. What changed in the last two days is where you can run it and what it costs you to try: Kilo's public gateway catalogue now carries a zero-priced twin of the model alongside the paid one, and StepFun's own account says the zero price lasts a week. Neither the vendor nor Kilo has published an end date, and Kilo's model data carries no expiry field at all. So the useful version of this story is not "there is a free week" — it is what the free entry actually declares, what it does not, and what the same work costs on the eighth day.

That distinction matters because Kilo is a coding agent pointed at a repository, and the difference between the free listing and the paid one is not only price.

What StepFun announced, and what Kilo's own data says

The promotion was posted from StepFun's account on Friday 9 October 2026: "Step 5 Preview is now free in @kilocode for one week." That is a dated statement of duration. Everything a reader can check afterwards is weaker than it.

Kilo publishes a machine-readable gateway catalogue, and it now contains two Step 5 Preview entries that differ in almost nothing but price:

• stepfun/step-5-preview-free — $0.00 per million input tokens, $0.00 per million output tokens, zero cache-read price, marked as a free model.

• stepfun/step-5-preview — $1.00 per million input tokens, $2.70 per million output tokens, $0.05 per million on cached input, marked as not free.

Both declare a 1M-token context window, a 64,000-token maximum output, and the same three reasoning-effort variants — low, medium and high, each exposed as a preset rather than a raw parameter.

What the free entry does not carry is a date. The catalogue's expiry field is empty for both entries, exactly as it is for every other model in the feed, and the free entry's creation timestamp is a 2025 value that the same feed also uses for other catalogue rows — a reused placeholder, so it cannot date the promotion. The paid Step 5 Preview row is more informative: its timestamp resolves to 8 October 2026, one day before the free twin appeared. Read together, that says Step 5 Preview arrived in Kilo's hosted gateway on 8 October and went free on 9 October. The week, and the fact that a week is what it is, comes from the post and from nowhere else.

One more gap is worth knowing before you go looking for the offer. Kilo's own curated Free Coding Models page — the page whose stated method is that it "updates from the same model data used by the app" — does not list Step 5 Preview. Its newest free listing is dated 2 October, and the models it does carry are Ling 3.1 Flash, Dots3-Note Preview, Laguna S 2.1 and the two Nemotron 3 entries. The app-facing model catalogue carries 156 models and does not include either Step 5 Preview row; the gateway feed carries 390 and includes both. So a reader who goes to the marketing surface will not find this promotion at all, and a reader who goes to the API will. That is not a contradiction so much as a latency: the free page's own copy warns that "availability can change when provider pricing changes or free promotions end," which is the honest thing for it to say given how promotions like this one arrive.

Two unrelated ways to reach Step 5 Preview inside Kilo Code

Kilo's onboarding copy and StepFun's onboarding copy describe the same product name and two different offers, and confusing them is the easiest mistake to make this week.

• StepFun's route — the vendor's documentation publishes a Kilo Code integration that runs through a Step Plan provider, with the base URL https://api.stepfun.ai/step_plan/v1 on the global platform or https://api.stepfun.com/step_plan/v1 on the China platform, an OpenAI-compatible Chat Completions path, and a Step API key you create yourself. Its stated prerequisite is "an active Step Plan subscription with available quota." StepFun sells that subscription in four monthly credit tiers, from a 400M-credit Mini plan up to a 40,000M-credit Max plan, and describes it as "the official harness for Step 5 Preview, with greater token savings."

• Kilo's route — the hosted gateway entries above, called with a Kilo account rather than a StepFun one, and free for the week. Kilo's pricing page puts free-model access on its "$0/mo, no hosted credit required" tier and notes that its Auto Free mode "routes to free models through Kilo Gateway where available."

So the free week is not a discount on StepFun's subscription and does not require a StepFun key. It is a hosted listing on somebody else's gateway, which is also why it can end without the vendor's pricing page changing at all.

The second difference between the two catalogue rows is the one to read carefully. The free entry declares a shorter capability list than the paid one. The paid row advertises reasoning_effort, response_format, structured_outputs, logprobs, top_logprobs, frequency_penalty and stop; the free row advertises max_tokens, temperature, tools, reasoning and include_reasoning, and nothing else. The paid row's modality string includes video input; the free row's reads text-plus-image. Those may be catalogue bookkeeping rather than real limits — the same model ID sits behind both — but a code path that depends on structured output or on video input is exactly the kind of thing to test on the free entry before you assume it behaves like the paid one.

The line that should decide whether you point it at work code

Every free model in Kilo's gateway feed carries the same flag set to true, and the paid Step 5 Preview row carries it set to false. The flag is mayTrainOnYourPrompts.

Read that plainly: on the free listing, Kilo's own catalogue tells you your prompts may be used for training. On the paid listing, at $1.00 and $2.70, it tells you they may not. Kilo's documentation treats this as a routing dimension rather than a footnote — its model-freedom material lists the policies you can filter on as training, prompt retention and provider location — which is a reasonable posture, but it does put the decision in your hands rather than in the default. The practical consequence is narrow and specific: a free week is excellent conditions for evaluating the model on public code, a scratch repository, or a task you would be happy to see in a training set, and it is the wrong conditions for a private client codebase, an unreleased product, or anything under an NDA, unless your organisation has already accepted that trade.

This is not a reason to skip the promotion. It is the reason the promotion is a choice rather than a free lunch, and it is the one fact about it that the announcement does not mention.

What the model costs on the eighth day

Step 5 Preview's list price is the number to plan against, because it is the one that survives the promotion. On StepFun's English pricing page the model bills at $1.00 per million input tokens on a cache miss, $0.10 on a cache hit, and $2.70 per million output tokens, with the caveat — printed on the page — that output tokens include the model's reasoning as well as its visible answer. Step 3.7 Flash, the smaller sibling that earlier Kilo promotions featured, sits at $0.20, $0.04 and $1.15 on the same three lines.

One discrepancy is worth flagging because it changes your cache maths. The vendor's Chinese pricing page lists the same model at ¥7 per million input tokens, ¥0.35 on a cache hit and ¥20 per million output — a 95% cache discount. The English page lists $0.10 on a cache hit against a $1.00 miss, which is a 90% discount. Kilo's gateway feed sides with the Chinese page, quoting $0.05 per million cached input. Two pages and a third-party catalogue, three readings of the same number; assume the worst of them when you model a cache-heavy agent loop, and check your own first invoice.

Then there is the number that makes the headline price misleading. Step 5 Preview is verbose. On Artificial Analysis's independent run of the model, completing the Intelligence Index suite consumed 160 million output tokens against a median of 82 million across the models scored in the same class — roughly double. At $2.70 per million, that verbosity is around $432 of the bill for a run whose total cost the same page puts at $1.03 per task. That arithmetic is the reason a free week is worth more here than for a lean model: for seven days you are not paying $2.70 a million to discover whether the model's reasoning traces are longer than you want.

What the free week is actually good for

The case for acting is straightforward. Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index, above the median of 26 for comparable models, at $1.00 and $2.70 — and the cost per index task, $1.03, is one of the better figures on that board because the model's speed offsets some of its verbosity. On paper it is a credible coding and long-context model at a competitive price. On your repository it is unmeasured. Seven days of hosted inference with a Kilo account and no per-token bill is the cheapest way to close that gap, and the tasks worth spending it on are the ones that are tedious to pay for: a full-repository review pass, a long-context refactor, a tool-calling harness you have never pointed at a 1M-token model, and a check of whether the reasoning-effort presets change the model's behaviour enough to matter for your workload.

The case for waiting is about trust, not capability. Nothing in this promotion is documented to a standard you can plan against: the duration exists in a post, the catalogue carries no expiry, and the curated free-model page has not caught up. If your workload needs a written end date before it can be scheduled, this is not that offer, and the honest advice is to treat the window as though it closes without warning and to keep anything load-bearing off it.

There is also a smaller point about how to spend the week well. The decision a free trial informs is never "is this model good" on its own; it is "is this better than what I am running, on my traffic, at a price I will actually pay." Half of that comparison is not free and does not need to be, because your incumbents are already callable. Which is the one place our own product genuinely belongs in this piece: OrcaRouter does not route Step 5 Preview — you can check that yourself against the public model list, where 205 models resolve and none of them are StepFun — so nothing here is a pitch to run it with us. What we do route is the comparison set you would hold it against, more than 200 models including Claude Opus 5.5, GPT-6.1 Sol, Kimi K3, GLM 5.3, DeepSeek V4.1 Flash and Gemini 3.8 Flash, behind one key and one bill, with provider list price passed through at 0% markup and automatic failover between providers. The reason that matters during a free week is switching cost: if the baseline lives behind one endpoint and the challenger only needs a base URL and a model ID changed, then evaluating Step 5 Preview on Friday costs you an afternoon, and abandoning it on the sixteenth costs you nothing.

The open question, and what settles it

Whether this is a one-off gesture or the beginning of a standing free tier in Kilo is genuinely unknown, and the two live readings point opposite ways. In favour of standing: Kilo runs a free-model programme with published methodology and a page that sorts free listings by freshness, and it already hosts free tiers of several open-weight models. Against: Step 5 Preview is a proprietary frontier model at $1.00 and $2.70, it is not open-weight, its free entry is absent from the curated free page, and one week is a short window for a model this expensive to serve.

What settles it is the catalogue, not the announcement. On the sixteenth of October there are three outcomes and they are distinguishable in public data. The free row still resolves at $0.00 with the same feature list — the offer stuck, or was renewed. The free row disappears and the paid row remains — the week was a promotion, and $1.00 and $2.70 is the price you plan against. Or the free row remains but gains an expiry field, which would be the first time this promotion had a written end date at all. Watching the gateway feed costs nothing, and it is a better instrument than waiting for a second post.