Crouton Alpha: A Free 1M-Token Model Nobody Will Put a Name On

The money is small. Take a coding-agent task that reads 200,000 tokens of repository and writes 20,000 tokens back: four cents on GLM-5.3-Flash, twenty-three cents on Gemini 3.8 Flash. Fill the entire 1,048,576-token window with one prompt and the input alone costs sixteen cents, or seventy-nine at Gemini's rate. That is the cash value of Crouton Alpha, free to call since it appeared on 3 November 2026 at 01:38 UTC, and it is not the interesting part.

The interesting part is that Crouton Alpha is a preview of something with no name on it. The provider field on its listing reads "Stealth"; the description says it was developed and operated by a third party "who has chosen to remain anonymous during this preview"; the routing platform is explicitly not its developer, owner or provider. There is no model card, no weights, no pricing schedule beyond zero, no benchmark, and no lab claiming it. What there is, is a specification: 1,048,576 tokens of context, text, image, video, file and audio input, mandatory reasoning, and a data clause that matters far more than the context number. This is what is known, what is checkable, and what a team should weigh before pointing an agent at it.

What the listing actually declares

Everything below comes from the preview's own record on 4 November 2026 — one day after it went up. None of it has been independently measured, and the specification is the vendor's claim about a vendor that will not say who it is.

• Identity — Crouton Alpha, listed 3 November 2026 at 01:38 UTC. Provider name "Stealth". Developer: unnamed.

• Context — 1,048,576 tokens of input. Maximum output 943,718 tokens, which is exactly 90% of the window, suggesting a derived default rather than a measured limit.

• Modalities — text, image, video, file and audio in; text out. Audio and arbitrary file input are the unusual part; most 1M-context coding models stop at text and image.

• Price — $0 input, $0 output. The record carries no expiration date, so the preview's end is not published anywhere.

• Reasoning — declared mandatory, with efforts from minimal to max and a default of max. It cannot be switched off.

• Tools — tool_choice "auto" only; structured outputs and response_format supported alongside temperature and top_p.

• Evidence — no benchmark table, no model card, no throughput or latency figure, tokenizer recorded as "Other", quantisation "unknown".

• Serving — one provider, no second endpoint, 100% uptime over the last 30 minutes, 5 minutes and day.

The description positions it as "a lightweight model built for scalable coding, agentic, and multimodal workflows", for "agentic sub-tasks, tool use, and document analysis in high-volume, repeated work". That is a description of what the model is for, written to be read by people who will route thousands of small calls through it. It says nothing about how good it is at any of that, and a day into the preview nobody outside the provider's walls has published a number.

The clause that matters more than the context window

Stealth listings carry a standard disclosure, and Crouton Alpha's is the one worth reading twice: prompts and completions may be retained by the provider but are not used for training, and "all other use" is governed by the preview's stealth terms. Both halves of that sentence are doing work. "Not used for training" is a promise about one specific downstream use. "May be retained" is permission for every other use, granted by a party you cannot identify and cannot audit, under terms written for a preview rather than a product.

Set that against what the model is built for. A coding model with 1M tokens of context and audio and file input is a model you feed a repository, a design document, a call recording, a PDF of contracts. That is precisely the material an organisation is least willing to hand to an unnamed third party for an indefinite period, and the free window is precisely the situation in which people are least careful. The listing also marks EU in-region routing as unavailable, so there is no option to pin processing to a European region while testing.

The quantisation field reads "unknown", which is a smaller point with the same shape: you cannot tell whether the model answering you is running at full precision, which makes any speed or quality impression you form during the preview a statement about this deployment, not about the model. If it ships later as open weights, your measurements may not transfer.

Reasoning is on, and it defaults to maximum

Two configuration details will cause most of the surprise, and both are mechanical rather than speculative. Reasoning is declared mandatory — an agent framework that expects to toggle thinking off cannot — and the default effort is max, with minimal, low, medium, high and xhigh available but not assumed. A client that sends no reasoning_effort gets the most expensive setting the model offers, on a model whose speed nobody has measured.

The other is tools. tool_choice accepts "auto" and nothing else: none, required and function are all unavailable. Frameworks that force a call with tool_choice:"required" — a common pattern when you want a guaranteed structured answer — will not get the behaviour they asked for. Structured outputs and response_format are supported, so the workaround exists, but it is a workaround. Combined with a 943,718-token output ceiling, the practical advice is unglamorous: set max_tokens explicitly, set reasoning_effort explicitly, and do not assume a free endpoint is a cheap one in wall-clock time.

Who is behind it, and what the last drops predict

Anonymous previews resolve, and the recent record is consistent enough to be useful. Ox Alpha appeared on 20 August 2026 with the same shape — free, 1M context, text and image in, positioned for coding and agentic work — and on 26 August Z.ai confirmed it as GLM-5.3-Flash, the 320B-total / 18B-active Mixture-of-Experts model now routed at $0.15 and $0.50 per million tokens. The community got there first, and not by guessing: published write-ups report counting prompt tokens against controlled probes, cancelling out the hidden chat template by differencing, and matching the vocabulary to GLM-5's on 95 of 95 probes with a mean absolute error of zero. Token counts, not vibes.

That episode is also a warning about the numbers that travel during a preview. A headline figure circulated for Ox Alpha that did not survive an actual run; the score that held up afterwards was far more ordinary. Stealth windows are where an impressive number can live for a week with nothing behind it, and the absence of any benchmark for Crouton Alpha is, on balance, the more honest configuration — there is no number to be disappointed by later.

For Crouton Alpha specifically, as of 4 November, there is nothing to go on beyond the specification. No tokenizer fingerprint has been published, no error-string comparison, no behaviour test. The listing's tokenizer field says "Other", so the vocabulary is not pre-matched to a known family, and the audio and file inputs widen the surface a fingerprint would have to explain. Anyone claiming on social media to know who built it, one day in, is guessing. The pattern from Ox Alpha suggests the answer lands within a week and usually only after the free period closes, which is a reason to watch rather than to speculate.

How to spend a free week without paying in something you would miss

There is a real use for a free window like this, and it is narrower than the announcement implies.

• Assume every prompt is published. Write the test set as if you were posting it in a public repo, because the clause permits retention and the terms cover everything beyond training.

• Keep it out of the production path. One provider, no second endpoint, no published expiry, and a model that can be withdrawn without notice is not a dependency; it is an experiment.

• Run the same prompts against a named model in the same session. An impression of a model with no baseline is worth nothing — the only useful output from the week is a diff, on your own tasks, against something you can price and keep.

• Pin the settings you care about. max_tokens and reasoning_effort on every call, structured outputs instead of forced tool calls, and log the request shape alongside the result so the comparison is fair.

• Log the failure modes. Rate limiting, timeouts and truncation are the things that decide whether a model is usable at volume, and a free preview is the cheapest place to discover them.

If you want the job done by a model with a name on it

OrcaRouter does not serve Crouton Alpha, and nothing here is a routing recommendation for it — the model exists only in that preview. The catalogue is where the same job gets done by something with a published vendor, a price and a second endpoint behind it. Prices below are from the OrcaRouter catalogue on 4 November 2026, passed through at the provider's list price with no markup.

• GLM-5.3-Flash — 1M context, text + image + video in, 128K max output, $0.15 in / $0.50 out per million. The model Ox Alpha turned out to be.

• DeepSeek V4.1 Flash — 1M context, text + image in, 384K max output, $0.15 / $0.60.

• Qwen3.8 Flash — 1M context, text + image + video in, 131K max output, $0.15 / $0.47.

• MiniMax M3.1 Flash — 1M context, text + image + video in, 524K max output, $0.15 / $0.60.

• Gemini 3.8 Flash — 1M context, text + image + file + audio + video in, 64K max output, $0.75 / $3.75. The closest match to Crouton Alpha's declared input set, and the reason it carries the arithmetic in this piece.

Four of those five sit at the same input price, which is the other reason the free window is worth so little in cash: long-context inference at the cheap end has become a commodity, and the decision that actually costs money is which one you standardise on. That is the decision a single API key is for — several named models behind one endpoint, automatic failover when a provider degrades, and a routing DSL for composing them, so testing a candidate and keeping production on a known model are the same integration rather than two. For the audio and file inputs specifically, Gemini 3.8 Flash declares the same five input types at the same context length, which makes it the honest comparison point for anything Crouton Alpha eventually ships as.

What would turn this into a launch story

Four things, in rough order of how much they would settle.

A lab claiming it. That is how Ox Alpha ended: Z.ai published the weights, the parameter count and the licence, and the mystery stopped being one. Weights would be the second — an open release would let anyone verify the context length and the modality claims in an afternoon, and would tell you whether the 1M window is native or stitched. A tokenizer fingerprint would come before either, since that is the method that has worked twice now and it needs no cooperation from the provider. And an independent benchmark, run on a known harness and published with the task set, would be the first evidence about quality this model has had.

Until one of those lands, Crouton Alpha is a specification with a retention clause attached, and the correct amount of confidence to place in it is the amount you would place in a listing.

If you have a bulk, repetitive, low-sensitivity job — parsing a shelf of public documents, classifying a corpus you would not mind seeing published — a free week on a 1M-token endpoint is a cheap way to find out whether long context helps your task at all, and the answer stays useful whichever model eventually turns out to be underneath. If the prompts are the asset, there is nothing here worth the trade: the saving is four cents a task, and the thing you would be handing over is the reason the task is worth running.

The distinction the card is drawing is the whole article in one line: every figure in the left column is a declaration by a provider that has not identified itself, and all six cells of the right column, the one that would carry verification, read as absences.

The comparison that is actually available

Because the subject cannot be benchmarked, the useful comparison is structural rather than numerical: what a named model at the same context length gives you that this one does not, and where the free window genuinely wins.

GLM-5.3-Flash is the closest thing to a control group: the model that the last anonymous 1M-context coding preview turned out to be, now with a vendor, a price, published weights and ten weeks in public behind it. If Crouton Alpha is a preview of something in that family — and there is no evidence either way yet — then the free week is a look at a model you will be able to rent for fifteen cents per million tokens afterwards, which makes the test cheap and the switching decision easy.

Run the numbers once and the shape of the decision stops being about money. A week of unlimited calls on a 1M-token model is worth cents to a few dollars to almost anyone; the reason to use it is information, and the reason to be careful with it is that the prompts you spend to get that information are worth considerably more than the tokens you save.