Crouton Alpha: Free API, Context Window and the Stealth Model Tradeoffs

Crouton Alpha became available through OpenRouter on November 3 as a free model for coding, tool use and multimodal agent work. It gives developers a million-token context window for experiments, but the company behind it remains anonymous and the free preview has no guaranteed lifetime.

The useful question is whether it can handle the repeated, well-defined steps inside an agent reliably enough to reduce the work left for a more expensive model.

Crouton Alpha API and context window

The OpenRouter listing identifies the model as stealth/crouton-alpha. It accepts text, images, audio, video and files such as PDFs, and produces text. Its context window is 1,048,576 tokens. Tool calling and JSON-schema structured output are supported.

OpenRouter describes it as a lightweight model for repeated coding, document analysis and agent subtasks. That is a statement about intended use, not a published parameter count or evidence that it beats a particular frontier model. OpenRouter routes requests to an anonymous third-party provider; it did not develop Crouton Alpha.

For an existing integration, the OpenRouter quickstart documents using an OpenAI-compatible client with the base URL https://openrouter.ai/api/v1. Select the exact model ID above rather than assuming that a :free suffix is required.

A sensible first test is a small, bounded task with an answer you can check: extract specified fields from a public document, produce a patch against a test fixture, or choose a tool from a restricted set. Then test the same task with longer inputs. A large advertised context window does not establish that every detail will be used correctly.

Free pricing has operational limits

Input and output tokens are currently priced at zero. Free access is a preview condition, not a dependable long-term cost assumption. OpenRouter’s Stealth Program terms allow a model to leave the program without advance notice.

The model page also says Crouton Alpha is unavailable through EU in-region routing. Teams that require that route should not interpret general API availability as satisfying their deployment requirements.

OpenRouter’s rate-limit documentation distinguishes platform limits from upstream capacity limits. Check the actual account and endpoint behavior rather than assigning Crouton a daily allowance from another free model. A free token price does not mean unlimited concurrent requests.

What happens to prompts and responses?

The model listing says the provider may retain prompts and completions but does not use them for training. No training is different from no retention. That distinction matters before connecting an agent to private repositories, internal documents or customer records.

The Stealth Program terms also restrict submitting sensitive information and make the anonymous provider part of the data flow. For an initial evaluation, public or synthetic material is the cleaner choice. A familiar routing service does not remove the need to understand the destination’s terms.

Early agent behavior needs testing

An OpenCode issue opened November 3 reports that Crouton Alpha stopped after a few tool calls and required another user prompt to continue. A second user added a similar report on November 4. The issue remains under triage; these reports do not establish a general failure rate or a confirmed cause.

They do identify a useful test that a single good answer misses. Does the model keep working until the requested result exists, or does it merely announce its next step and stop?

For a coding workflow, measure completed tasks, passing tests, retries and human interventions alongside token cost. Keep permissions narrow and inspect changes before allowing them to affect real systems. If a free model needs frequent nudges or another model to repair its work, that overhead belongs in the comparison.

Crouton Alpha makes it easier to test a division of labor between a stronger planner and a cheaper executor. The deciding evidence will be how much verified work that executor finishes without supervision, and whether that remains useful when the preview ends.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help AI founders turn technical launches and product decisions into clear stories developers can use and share.

Book a 30-minute call.