OpenRouter
stealth/crouton-alpha
From Europe Every request failed with a provider error before the first token.
Elsewhere $0 per million tokens in and out. Needs an OpenRouter account.
Open the OpenRouter listing ↗STEALTH MODEL · FREE PREVIEW · SINCE 3 NOV 2026
Crouton Alpha
A free reasoning model in a crouton costume. Nobody has owned up to baking it.
Blocked in Europeon OpenRouter and OpenCode
Named in French.
Blocked in France.
- 1M
- tokens of context
- $0
- input and output
- 5 inputs
- text, images, audio, video, files
- 6 levels
- of reasoning, always on
Where are you?
01 · CAN YOU USE IT?
Blocked in Europe on both routes.
We tried from Europe on 3 November. Neither route let us in. Outside Europe it’s free, though some other countries are blocked too.
OpenCode
opencode/crouton-alpha-free
From Europe “This model is not available in your country.”
Elsewhere Free in the OpenCode app for one week from 3 November.
Open OpenCode’s model list ↗FIRST DAY ON OPENROUTER
- 80M+
- tokens used on launch day
- Kilo Code
- its busiest app, ahead of ZCode and Chub AI
- 13.7%
- availability: most requests got no answer
02 · INGREDIENTS
What’s in the crouton.
OpenRouter pitches it as “a highly efficient reasoning model built to scale out the work your agents repeat all day.” The label has the rest.
Crouton Alpha
Serving size: one request
Context window1,048,576 tokens
Longest answer (tokens)943,718 on OpenRouter
270,000 on OpenCode
270,000 on OpenCode
ReadsText, images, audio, video, files
WritesText
ReasoningAlways on, six levels from minimal to max. Shows a summary, hides the rest
ToolsFunction calls and structured output
Price$0 in and out
Your promptsMay be kept by the provider, not used for training
Contains: bread. May contain: Meta.
03 · WHO BAKED IT?
Probably Meta. Not a Muse we know.
Nobody official has said who made Crouton Alpha. Four clues point to Meta, the company behind the Muse models. One says it isn’t a Muse model we’ve met.
-
Six levels, one family.
You can tell Crouton how hard to think, on six levels from minimal to max. We compared that menu across all 454 models on OpenRouter. Only Meta’s Muse Wonder and Muse Spark 1.3 offer exactly the same six.
So: a menu that rare looks like a family trait.
-
Same million, same ears.
It can take in 1,048,576 tokens at once, exactly as many as both Muse models. And like Muse Wonder, it listens to audio as well as reading images and video.
So: it fits, though other big models share that size too.
-
Same borders, same blind spot.
Within minutes of the launch, the model tracker @aitrackerbot noticed that OpenCode blocks the same countries as Meta’s own model service, and that Crouton trips over a rare Japanese word that every Muse model they tried also gets wrong.
So: two more family traits.
-
It mutters like Muse.
Crouton never shows its full thinking, only a summary of it: one short line per step, each starting with an “-ing” word, run together. Muse models write theirs the same way, down to a stock phrase.
Crouton Alpha Generating an SVG beach sunset scene with animals, birds, and reflections under abstract output constraints.
Muse Wonder Generating a detailed Mona Lisa SVG with paths, gradients, and filters for sfumato under abstract output constraints.
Crouton Alpha Resolving camera basis and perspective projection…
Muse Spark 1.3 Evaluating camera frustum coverage and occlusion…
So: same habits, most likely the same kitchen.
-
But it counts like a stranger.
Models read text in tokens, small pieces of words. Each family cuts text up its own way, so the count for the same phrase works like a fingerprint. We had Crouton and two Muse models count 24 test phrases. The two Muse models agreed on all 24. Crouton agreed with them on 2.
It matches none of the 42 other models we have fingerprinted either. Even its closest match is less alike than two unrelated models usually are. Numbers, sums, dates and emoji cost it about half what they cost every other model.
So: not a Muse model we know, served as it is. A Meta model with a new way of cutting up text would still fit, and so would a provider that counts its own way.
CLUE 6 · THE SCORES
Sitting with the family.
Crouton’s dot beside the Muse models and other fast models.
What the evidence says: a new model from Meta, but not a Muse model we know.
STILL OPEN
- Who made it. No official reveal yet.
- Which countries are blocked. No official list. We only know about Europe, from our own tries.
- Why it counts differently. A new way of cutting up text, or a provider that counts tokens its own way.
- How much it really thinks. Only the provider knows. Our estimate is next.
04 · HIDDEN THINKING
It thinks far more than it admits.
AI models read and write in tokens, small pieces of text, usually part of a word. Thinking costs tokens too, and Crouton does most of its thinking out of sight.
What you see
Crouton always thinks before it answers, but keeps that thinking to itself. It shows a summary instead, a short line per step, like “Resolving camera basis and perspective projection…”
What it reports
With every answer, the provider also reports how many tokens of thinking went into it. For Crouton, that count only covers the summary lines.
What the clock says
Yet its answers to our benchmark tasks took minutes, and Crouton writes fast. Minutes of writing at that pace add up to far more than a few summary lines.
What we counted
So we timed each answer, multiplied the time by its writing speed and took away the answer you can read. What is left is the thinking it did out of sight.
more thinking than it reports
Locked in a box. In our game runs on 4 November, Crouton also sent its thinking back encrypted, so we can’t read it. For a single game move, that sealed block ran to about 233,000 characters.
It grows with the task. The portrait needed little hidden thinking, the physics scenes the most. The total lands close to what Muse Wonder openly reports for the same tasks.
Why it matters. Our efficiency scores weigh how good an answer is against how much thinking it took. Taken at its word, Crouton would look as if it barely thinks. For models that report only their summaries, the scores use our estimate instead.
Advertisement · DEMODOKOS FOUNDRY
This one works in Europe.
Demodokos Foundry makes music and expressive voices on your own Windows PC, wherever that PC is. France included.
TODAY’S SPECIAL
Crouton in the Soup
Two voices and music · made in Foundry0:00 / 0:53
The crouton needs to talk.
05 · STEALTHMARK
How Crouton scores.
StealthMark is our score for hard benchmark tasks, split into general and visual intelligence.
Today’s tasting menu
06 · QUICK ANSWERS
Crouton Alpha, briefly.
Can I use Crouton Alpha in Europe?
No. OpenRouter and OpenCode both turned us away from Europe on 3 November 2026. France too, despite the name.
Is Crouton Alpha free?
Yes, during the preview. OpenRouter charges $0 for input and output, and OpenCode offers it free for one week from 3 November.
Who made Crouton Alpha?
Nobody official has said. Its reasoning levels, the way it summarises its thinking and a community test of its blocked countries all point to Meta. But it counts text differently from every Muse model we tested, so it looks like a new Meta model rather than a Muse we know.
How much does Crouton Alpha really think?
Far more than it reports. The provider counts only its short thinking summaries, while its answers take minutes. Our estimate from time and writing speed is in the hidden thinking section.
What can Crouton Alpha read?
Text, images, audio, video and files, with a context of 1,048,576 tokens. It answers in text and always reasons first.
Is Crouton Alpha good?
Yes. It holds its own on our hard benchmark tasks. The scores are in the StealthMark section.