Crouton Alpha: Free Stealth Model, 1M Context, 30-Day Logs

Crouton Alpha is a model nobody will name. It has a 1,048,576-token context window, takes text, images, video, PDFs and audio, costs nothing this week, and the company running it has asked to stay anonymous. Two gateways, OpenCode and OpenRouter, are routing traffic to it.

OpenCode is a coding-agent tool, and OpenRouter is a marketplace that sells access to many models through one API. A "stealth model" is a pre-release model that a provider tests under a codename. Here's what the documentation actually says before you paste a repository into it.

OpenCode's announcement on X, Nov. 3, 2026 (@opencode)

View on X

What the listings confirm

OpenRouter's model page shows a release date of Nov. 3, 2026, a free price for prompts and completions, and a "1.0M" context, which its FAQ gives as 1,048,576 tokens. The listing says it accepts "text, images, video, files such as PDFs, and audio as input and returns text." Output is text only.

OpenRouter's public model list adds a few details the page doesn't. The entry shows a maximum completion of 943,718 tokens, support for tool calling, structured outputs and a reasoning-effort parameter, and an empty expiration date. We checked that list directly, so those fields are what the API reports, not a benchmark of the model.

On OpenCode, the console's models page lists it as crouton-alpha-free with a $0 price for input, output and cached reads. The page says the team is using the free period "to collect feedback and improve the model."

Who runs it, and who doesn't

OpenRouter is explicit: the model "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview," and OpenRouter "routes requests to it and is not its developer, owner, or provider." OpenCode's page doesn't name a developer either.

OpenRouter describes it as a lightweight model for coding, agent sub-tasks, tool use and document analysis. That's the provider's pitch. We found no benchmark, speed figure or independent test in any listing we checked, so how well it codes is unknown.

"No training" is not "no logs"

The announcement says "No training on data." That part holds up, but the two gateways describe it differently. OpenCode says that during the free period, collected data "is retained for 30 days and is not used for model training." OpenRouter says prompts and completions "may be retained by the provider but are not used for training."

The same OpenCode page says its default is zero retention, then lists exceptions, and Crouton is one of them. Reading down that list is the most useful thing in this story:

• Crouton Alpha Free: kept 30 days, not used for training.
• Pulse Alpha Free, another stealth model: may be retained and used to improve the model.
• Flint 40B Free: may be used to improve the model, kept 30 days.
• Beam Free: not used for training, kept 30 days for safety.
• GLM-5.3-Flash Free and DeepSeek V4.1 Flash Free: zero retention, no training, according to the page.
• Nemotron free endpoints: trial use only; the page says not to submit personal or confidential data.

Our reading: Crouton has the stricter terms among the stealth entries, but a 30-day log held by a company you can't identify is still a log. Keep proprietary code, customer data and credentials out of it, and send public or throwaway material.

Where you can use it

The OpenCode page says it isn't available in the European Union, China (including Hong Kong and Macau), Russia, Iran, North Korea, Cuba, Venezuela, Pakistan and a longer list of other countries, "or in some regions within otherwise supported countries." The United States isn't on the list, so it should work for US users, though the subregion wording leaves room for exceptions. OpenRouter's page separately says the model isn't available through its EU in-region routing.

The free week has no printed end date

OpenCode's post says "free for the next week." Its docs say "for a limited time" and give no date, and OpenRouter's API lists no expiration. Counting a week from Nov. 3 gives Nov. 10, but that's our arithmetic. One aggregator site gives the same date; we didn't treat it as a source.

Rate limits are the other gap. OpenCode's page lists none for this model. OpenRouter's limits documentation covers its :free model variants: 20 requests per minute, and 50 requests per day for accounts that bought less than $10 of credits all-time, or 1,000 per day for accounts that bought at least $10. If you're testing a long-context job through OpenRouter's free route, 50 requests is a small day.

A week is enough to test one thing

Pick one job where a million tokens would change the answer, such as a whole repository, a long contract set or an hour of video, and run it with non-sensitive material. Record whether it finishes and what it gets wrong. When the free price disappears, you'll have a result, which is more than a stealth model's own listing gives you.

Source: OpenCode

crouton alpha opencode openrouter stealth model free ai models

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.