Skip to content
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog

What Is Crouton Alpha? OpenRouter's Stealth Model, Graded

At a glanceQuick answers
What is Crouton Alpha?
A free stealth model on OpenRouter since November 3, 2026: 1,048,576-token context, text, image, video, file and audio input, built for repeated agent sub-tasks.
Who made Crouton Alpha?
Unconfirmed. Early clues point to Meta’s Muse family, but its token counts do not match Meta’s, and no company has claimed it.
Is Crouton Alpha free?
Yes, $0 in and out during the preview. Every earlier OpenRouter stealth model lost its free price once its maker was named.
Editorial data illustration on a near-white ground: a sealed teal crate stamped Crouton Alpha with a question mark, beside a huge navy number 1M for its context window, and three small clue cards pinned to the crate with one amber card set apart
Fig 0A free model, a million tokens, and no name yet. Made by CellCog's image agent, running GPT Image 2.5.

OpenRouter put a new anonymous model in front of developers on November 3, 2026. It is called Crouton Alpha, it is free, it holds a million tokens and it takes almost any input. OpenRouter’s post on X called it “A highly efficient reasoning model built to scale out the work your agents repeat all day.”

Nobody has claimed it. Within 20 minutes of the announcement, the first fingerprints were pointing at Meta, with one clue that does not fit. Here is what OpenRouter published, what the clues show, and how earlier stealth models were unmasked.

Key points5 · 6 min full read
  1. A sealed crate with a question mark: a model with no named maker.
    OpenRouter announced Crouton Alpha, a free stealth model, on November 3, 2026 at 20:17 UTC; its catalog entry is timestamped 01:38 UTC the same day.
  2. Five small input icons flowing into one box: many kinds of input, text out.
    It takes text, images, video, files and audio, holds 1,048,576 tokens of context, can write up to 943,718 output tokens, and costs $0 in and out.
  3. A filing cabinet with a crossed-out graduation cap: kept, not trained on.
    OpenRouter says the provider is anonymous, and that prompts and completions may be retained by the provider but are not used for training.
  4. A magnifying glass over three clue cards: the evidence so far.
    Within 20 minutes a tracking account pointed at Meta: the same blocked-country list as Meta’s model API, a Japanese word every Muse model fails on, and Muse-only effort settings.
  5. One amber card set apart from a stack: the clue that does not fit.
    Not every clue fits: its token counts do not match Meta’s, by the same account’s own test. No maker has claimed it.

§ 01What OpenRouter published

The listing says the model “is developed and operated by a third-party provider who has chosen to remain anonymous during this preview.” OpenRouter “routes requests to it and is not its developer, owner, or provider.”

Item Value
Model id stealth/crouton-alpha
Catalog entry November 3, 2026, 01:38:50 UTC (created timestamp)
Announced November 3, 2026, 20:17 UTC (OpenRouter on X)
Maker Anonymous third-party provider
Context and output 1,048,576 tokens in, up to 943,718 out
Input Text, image, video, file and audio; text out
Reasoning Adjustable, with reasoning effort settings
Tools Tool calling with tool_choice auto only, and structured output
Price $0 input, $0 output
Data “Prompts and completions may be retained by the provider but are not used for training”
Table 1Crouton Alpha, the OpenRouter listing (read November 3, 2026, 20:36 UTC)

The description pitches it for “executing agentic sub-tasks, tool use, and document analysis in high-volume, repeated work.” That is the job a fast, cheap model does under a bigger planner: the repeated steps, not the plan.

§ 02The clues, graded

The account AI Tracker Bot posted the first identification at 20:29 UTC: “Crouton Alpha looks like Meta”. Its evidence, and our own read of OpenRouter’s catalog:

Clue Source Points to Our read
Same 21 blocked countries as Meta’s model API AI Tracker Bot Meta Specific, not checked by us
Fails on a rare Japanese word every Muse model fails on; 93 other models do not AI Tracker Bot Meta Strong if it holds up
Only Meta’s Muse listings share its effort settings AI Tracker Bot Meta Fits OpenRouter’s catalog
tool_choice limited to auto AI Tracker Bot, OpenRouter catalog Meta, per the account OpenRouter’s metadata for Meta’s own Muse Spark 1.3 lists all four options
Tokenizer labeled Other OpenRouter catalog Meta, not Google Every Gemini listing says Gemini; Muse listings say Other
Token counts do not match Meta’s AI Tracker Bot Against Meta The one clue that does not fit
Table 2The identity clues, graded (November 3, 2026, 20:36 UTC)

The account itself says the match is not confirmed. Its follow-up added the tool_choice point and then the problem: “its token counts don’t match Meta’s“. Token counts are usually the strongest fingerprint, which is why one mismatch matters.

Animation: three teal clue cards, Blocked countries, Japanese word test and Effort settings, rise from a sealed Crouton Alpha crate and drop into a Meta Muse box with check marks; an amber Token counts card stops short with a grey cross, the Google Gemini box stays empty, and Not confirmed appears under the crate
Fig 1Three clues point to Meta Muse; token counts do not fit; nobody has claimed it

The size limits fit more than one family. In OpenRouter’s catalog, Crouton’s 943,718 output tokens match Meta’s Muse Spark listings and Google’s Gemini 4 Argon, while Gemini 3.8 Flash stops at 65,536.

Maximum output tokens on OpenRouter, thousands (read November 3, 2026)Bar chart of maximum output tokens in thousands: Crouton Alpha 943.7 highlighted, Meta Muse Spark 1.3 943.7, Gemini 4 Argon 943.7, Gemini 3.8 Flash 65.5Crouton Alpha943.7Meta Muse Spark 1.3943.7Gemini 4 Argon943.7Gemini 3.8 Flash65.5Maximum output tokens on OpenRouter, thousands (read November 3, 2026)Bar chart of maximum output tokens in thousands: Crouton Alpha 943.7 highlighted, Meta Muse Spark 1.3 943.7, Gemini 4 Argon 943.7, Gemini 3.8 Flash 65.5Crouton Alpha943.7Meta Muse Spark 1.3943.7Gemini 4 Argon943.7Gemini 3.8 Flash65.5
Fig 2Maximum output tokens on OpenRouter, thousands (read November 3, 2026)

So the limits rule out Gemini 3.8 Flash but not Meta or a larger Google model. The tokenizer label and the three tracker clues lean Meta; the token counts do not.

§ 03The stealth record

OpenRouter has named every earlier stealth model it tracks here, and the free price ended each time.

Days each OpenRouter stealth model stayed anonymousBar chart of days anonymous: Quasar Alpha 11 (GPT-4.1 test), Horizon Alpha 8 (GPT-5), Pony Alpha 5 (GLM-5), Ox Alpha 6 (GLM-5.3-Flash), Union Alpha 1.4 (Pareto by Unbiased) highlighted, Space Bunny Alpha 25.8 (MiniMax M3.1-Flash)Quasar Alpha11Horizon Alpha8Pony Alpha5Ox Alpha6Union Alpha1.4Space Bunny Alpha25.8Days each OpenRouter stealth model stayed anonymousBar chart of days anonymous: Quasar Alpha 11 (GPT-4.1 test), Horizon Alpha 8 (GPT-5), Pony Alpha 5 (GLM-5), Ox Alpha 6 (GLM-5.3-Flash), Union Alpha 1.4 (Pareto by Unbiased) highlighted, Space Bunny Alpha 25.8 (MiniMax M3.1-Flash)Quasar Alpha11Horizon Alpha8Pony Alpha5Ox Alpha6Union Alpha1.4Space Bunny Alpha25.8
Fig 3Days each OpenRouter stealth model stayed anonymous

They were revealed as an early GPT-4.1 test (Quasar, April 2025), an early GPT-5 checkpoint (Horizon, July 2025), Z.ai GLM-5 (Pony, February 2026), Z.ai GLM-5.3-Flash (Ox, August 2026), Pareto by Unbiased (Union, September 2026) and MiniMax M3.1-Flash (Space Bunny, September 2026).

Fingerprints were right on Ox Alpha and Space Bunny Alpha, and wrong on Union Alpha, a blended model from a startup nobody guessed. A clean match on three clues and a miss on token counts is closer to the Union case than the Ox one, so we grade Meta as likely, not settled.

§ 04What it means

For builders, a free, fast model with a million tokens is worth testing on the repeated steps of an agent: extraction, tagging, short tool calls. Keep sensitive data out of it: the provider may keep what you send, and you cannot name the provider.

Our conflict, declared: CellCog’s AI employees run on Claude Opus 5.5, and we do not use Crouton Alpha. We have submitted a CellCog connector to Meta’s Muse directory, so read the Meta clues with that in mind.

§ 05Sources

Frequently asked5 questions

Q1What is Crouton Alpha on OpenRouter?

A stealth model listed as stealth/crouton-alpha on November 3, 2026. OpenRouter describes it as a lightweight model for scalable coding, agentic and multimodal workflows, and says an anonymous third-party provider runs it.

Q2Is Crouton Alpha made by Meta?

Not confirmed. A tracking account found three Meta-shaped clues within 20 minutes of the announcement, and posted that its token counts do not match Meta’s. No company has named it.

Q3Is Crouton Alpha Google Gemini?

Its 1M context and audio and video input also fit Gemini models. But OpenRouter labels its tokenizer Other, as it does for Meta’s Muse listings, while every Gemini listing says Gemini.

Q4What happens to my prompts?

OpenRouter’s page says prompts and completions may be retained by the provider but are not used for training, with other use under its Stealth Model Terms. Treat anything you send as readable by a company you cannot name.

Q5How long will Crouton Alpha stay free?

Nobody has said. Earlier OpenRouter stealth models stayed anonymous between 1.4 and 25.8 days, and the free price ended when each maker was named.

Published 03 November 2026 All Choosing a platform →