Step 5 Preview Is Free in Three Places This Week — and That Is the Story

Step 5 Preview has been made free to try in three separate developer surfaces inside five days — the Cline extension on 8 October, the Nous Portal from 8 October to 15 October, and Kilo Code from 9 October. The model is not new: StepFun announced it on 20 September 2026 and opened its own API and Studio the same day at $1.00 per million input tokens and $2.70 per million output tokens. So the thing worth writing about this week is not the model, which has been callable for three weeks, but the campaign trail around it. Three free runs in five days, from three different partners, on a model whose weights were promised for 15 October, is a distribution strategy you can read — and it tells you more about what happens to Step 5 Preview after this week than any single free-week announcement does.

This piece is about the shape of that campaign: what is genuinely free and where, what the free access does not change, and how to spend the week so that the evidence you generate is still useful in November. Every date and price below is attributed to whoever published it, because the three surfaces are three different companies with three different sets of terms, and none of them is StepFun's own platform.

What is free, where, and for how long

The campaigns are documented to different degrees, and the differences matter when you are deciding whether to move a real workload onto one of them.

• Cline — announced by StepFun on 8 October 2026 as a one-week free run of Step 5 Preview inside the Cline coding agent. The window is stated only in prose ("one week"); no end date is published anywhere we could find.

• Nous Portal — listed at $0.00 per million tokens for Step 5 Preview from 8 October through 15 October 2026, alongside StepFun's smaller Step-3.7-Flash at the same $0.00. This is the only one of the three with an explicit calendar window, and it is the one StepFun's own post of 9 October pointed at.

• Kilo Code — announced by StepFun on 9 October 2026 as free "for one week", and reflected in Kilo's public gateway data as a zero-priced row for the model. As with Cline, the length is stated in the announcement and the expiry is not published in the data feed.

The pattern is the same in all three cases: a partner absorbs the token cost for a short window, and the vendor gets usage data, integration work from the partner's engineering team, and a stream of builder reports. The free access is real — the same model that bills at $1.00 and $2.70 per million tokens on StepFun's platform is listed at zero on those surfaces — but it is a marketing window, not a pricing change.

Two things deserve saying plainly. First, none of these surfaces is ours; we name them because they are the news, and naming a campaign is not a recommendation. Second, the terms you accept when you use one of them are that partner's terms, not StepFun's.

What the free week does not change

The list price is the one number the campaigns do not move. Step 5 Preview carries a published rate of $1.00 per million input tokens and $2.70 per million output tokens on the vendor's own platform, per StepFun, and that is what it reverts to on 16 October on the surfaces where the window is published. If your evaluation ends with a decision to adopt, the adoption decision is priced at list, not at zero.

The weights are the other moving part, and they have moved in a direction nobody announced. StepFun's launch materials promised BF16 weights on 15 October 2026; the vendor's own Hugging Face repository for that build was a stub at announcement time — one .gitattributes and nothing else — and a check from this desk today returns an authentication error rather than a file listing, which is what the platform serves for a gated or unfilled repository. Meanwhile a community GGUF build of the model, published on 2 October and last modified on 6 October, has been moving: its metadata reports roughly 601 billion total parameters, a 1,048,576-token context length and about 275 GB of quantised files, and it has been downloaded around two thousand times in eight days. That is a third-party conversion, not a vendor release, and nobody outside StepFun has vouched for its fidelity — but it is evidence that weights for this model exist in distributable form while the official repository for them does not. That makes 15 October a soft date with a hard fallback: the model may be self-hostable before the vendor says it is, with all the caveats that implies about provenance and quantisation quality.

Third, the free listing does not change the model's economics, because most of the cost of a verbose reasoning model is output tokens and output tokens are exactly what these campaigns let you spend freely. That is the honest reading of a free week on a model whose most-cited efficiency criticism is verbosity: the week is cheap for you and informative for the vendor, and the evaluation you run should be designed to answer a question that survives the price going back to $2.70.

Fourth, and worth stating because it comes up in every one of these campaigns: Step 5 Preview is not available through OrcaRouter. We do not route it, and no amount of free access elsewhere changes that. What we do route is a catalogue of more than 200 models — Kimi K3, MiniMax M3, the DeepSeek and Qwen families, the Claude and Gemini lines and a great deal else — behind one API key and one bill, at 0% markup with provider list price passed through. That is the substrate you would use to compare Step 5 Preview against the models you already trust, which is the comparison that decides anything.

Three campaigns in five days is the signal

Launches in this market are usually one announcement on one platform with one partner. Three free integrations in five days is a different move, and it is the second time StepFun has run this play: the same model was free in Cline on 8 October, in the Nous Portal starting the same day, and in Kilo Code the next day. Whatever the internal reasoning, the observable effect is that Step 5 Preview is being distributed through coding agents and aggregation portals rather than through a single flagship integration, with each partner seeded separately.

There is a practical consequence for the reader. Availability of this model is now a compatibility question rather than a pricing question: if you already live inside one of those three tools, you can try a 600-billion-parameter flagship today without a purchase order, and if you do not, the free week passes you by entirely. That is a real asymmetry between teams, and it is worth knowing before the deadline rather than after it.

There is also a signal in the timing. The vendor's own launch material put the open-weights build on 15 October, the same day the Nous Portal window closes. Free access across three partners in the fortnight before a promised weight drop is what pre-adoption looks like: usage, integration and community evidence are being accumulated ahead of the release that will let teams self-host. If the weights ship on schedule, the free week will look in hindsight like the on-ramp. If they do not, the free week will have been the only thing that changed in October, and the model's story is unchanged from 20 September.

How to spend the week so the evidence still matters later

A free week is only worth something if it produces an answer you can act on once the price returns. Three things are worth measuring, and none of them requires paying for tokens.

First, measure it against the model you already run, on your own tasks, with the same harness — not against a published leaderboard. The independent board puts Step 5 Preview at 44 on the Artificial Analysis Intelligence Index, well above the 26 median among comparable models, and measured its output at 87 tokens per second against a 78 average, at $1.03 per index task. Those are the strongest third-party facts available, and they still cannot tell you whether it writes your migration scripts better than the model you have in production.

Second, measure the verbosity directly, because that is where the list price lands. On the same evaluation the model generated about 160 million output tokens against a median of 82 million, which at $2.70 per million is a real multiple on output spend. If your workload is long structured output, that number will matter more to your bill than any per-token discount. If it is short tool-call turns, it will matter much less.

Third, write the decision down before the window closes — what you tested, what the failure modes were, and what it would have to cost for you to move. A model that is free on Tuesday and $2.70 per million output on the following Thursday is the same model, and the only thing you gained in the free window is evidence. Evidence is worth keeping; a nice week of free tokens is not.

And when the window shuts, the comparison does not have to shut with it. OrcaRouter passes provider list prices through at 0% markup, so a vendor price cut for a model we do carry is live on your key the same day, and automatic failover means a promising-but-unproven model never has to carry a production path alone while you decide. Step 5 Preview is not behind that API today. The models you would judge it against are.

What to take from this week

Step 5 Preview is a released model with a published price, a 1M-token context, a 600B/27B mixture-of-experts layout per StepFun, and one independent score of 44. This week it is free in three places with three different sets of terms and two different published end dates, and next week at least one of them is over. Treat the free access as a trial you were handed rather than a price you were given, run the comparison you actually care about, and price the outcome at $1.00 and $2.70 per million tokens, because that is what the model costs the day the campaigns end. The open weights are the other date on the calendar, and until the repository has files in it, that one is still a promise.

We do not route Step 5 Preview, but the models you would compare it against sit behind 0% markup with provider list price passed through , with automatic failover keeping a production path alive while you evaluate.