A free reasoning model that reads 1 million tokens
OpenRouter, a U.S. AI model routing service, released "Crouton Alpha," a "stealth" model whose developer is not disclosed, for free on Nov. 3 (Nov. 4 Japan time). It reads 1 million tokens (units of text) at a time and accepts images, audio and video in addition to text. The company described it as a highly efficient reasoning model built to scale out the work that agents (systems in which AI carries out tasks on its own) repeat all day.
🥷 New stealth model: Crouton Alpha A highly efficient reasoning model built to scale out the work your agents repeat all day. - Free to use - 1M token context - Text, image, audio, and video input - No training on data
According to OpenRouter's model list, the context window is 1,048,576 tokens and the maximum for a single response is 943,718 tokens. It supports tool calling and output in a fixed format. The listing describes it as a lightweight model for handling large volumes of programming, agent subtasks and document analysis, and suggests pairing it with a stronger reasoning model.
3 reasons to think it is made by Meta
About 10 minutes after the announcement, @aitrackerbot, an X account that tracks model appearances, posted that Crouton Alpha looks like a Meta model. It gave three reasons. The 21 countries where it is unavailable, listed on the model's page on the AI coding tool OpenCode, matched those of Meta's model API. It made the same error on a rare Japanese word that all of Meta's "Muse" family of models get wrong. None of 93 other models got it wrong, the account said. Of 454 models on OpenRouter, only Meta's Muse matched its options for reasoning effort settings. The account added, "Not confirmed."
Well hello, new stealth model on OpenRouter! Crouton Alpha looks like Meta 🍞 - OpenCode's page for it lists the same 21 blocked countries as Meta's model API, down to Eritrea, Haiti and Western Sahara - It breaks on a rare Japanese word that every Meta Muse model breaks on. 93 other models: 0 misses - Of 454 models on OpenRouter, only Meta's Muse listings share its effort settings Not confirmed.
According to CellCog, the 943,718-token maximum for a single response is the same as that of "Muse Spark 1.3," which Meta released in September. The tokenizer is also displayed as "Other," the same as Muse.
Some clues do not fit
AI service CellCog said it is likely but not confirmed that the model is made by Meta. According to CellCog, the detecting account said the token counts for text do not match Meta's models. On how tool calling is specified, Muse Spark 1.3 accepts four options, while Crouton Alpha accepts only "auto." The maximum response value is also the same as Google's "Gemini 4 Argon," so the possibility of a large Google model cannot be ruled out, it said. CellCog disclosed in the article that it has an interest, having applied for its own integration with Meta's service.
OpenRouter says the provider wants to remain anonymous during testing. Meta has not made an announcement.
Past anonymous models identified within 1 to 26 days
OpenRouter has released models with undisclosed developers for free before. According to CellCog's tally, "Quasar Alpha," an early version of OpenAI's "GPT-4.1," was revealed after 11 days, and "Space Bunny Alpha," which was "M3.1 Flash" from China's MiniMax, after 25.8 days. Outside analyses were correct for "Ox Alpha" from China's Zhipu (Z.ai) and for Space Bunny, and wrong for "Union Alpha" from Unbiased of the U.S.
The day before, on Nov. 2, "Pulse Alpha," a free model whose developer was not disclosed, also appeared on OpenCode. Free models with hidden developers are appearing one after another on routing services and AI coding tools.
Meta pushes low-priced API
If it is made by Meta, it would expand the lineup of the developer API the company launched in July. Meta launched the "Meta Model API" as a public preview on July 9, offering "Muse Spark 1.1" at $1.25 for input and $4.25 for output per million tokens. Muse Spark 1.3, released in September, has the same pricing on OpenRouter, and the higher-end "Muse Wonder," released in October, costs $4 for input and $20 for output.
In its October earnings, Meta put its 2026 capital spending forecast at $135 billion to $145 billion, raising the low end by $5 billion. Capital spending in the July-September quarter was $34.9 billion, exceeding the $32.8 billion in cash generated from operating activities. Whether the company can recoup its huge AI investment through outside sales has become a focus for investors.
- Muse Spark 1.3
- Muse Wonder
Free, but inputs may be stored
Crouton Alpha can be used on OpenRouter by specifying the model name "stealth/crouton-alpha," with both input and output at $0. No free period has been given, and under OpenRouter's terms, stealth models are offered for a limited time and can be withdrawn at any time. Free models have limits on the number of requests per minute and per day. A setting to process data only within Europe is not supported.
OpenRouter says data is not used for training. However, inputs and responses may be stored by the provider, and under the terms for stealth models, content may be shared with the anonymous provider. Care is needed in sending confidential materials or personal information while the developer remains unknown.
- Meta releases Muse Spark 1.3Official
- Meta releases higher-end Muse WonderOfficial
- Crouton Alpha registered in OpenRouter's model listReport
- OpenRouter releases Crouton Alpha for freeOfficial
- @aitrackerbot posts analysis that it is made by MetaRumor
