Crouton Alpha

Crouton Alpha is an anonymous stealth reasoning model listed by OpenRouter on 2026-11-03 as stealth/crouton-alpha. OpenRouter describes it as a lightweight model for scalable coding, agentic, and multimodal workflows—especially high-volume agent sub-tasks, tool use, and document analysis—often paired with frontier reasoners for the bounded steps that turn plans into action.

Listed capabilities: about 1,048,576 tokens of context; text, image, audio, video, and file (e.g. PDF) inputs with text output; tool calling and structured outputs; $0 input/output during the preview. A third-party provider stays anonymous while OpenRouter routes traffic; prompts/completions may be retained but are said not to be used for training. It also appeared on OpenCode and Perch AI; community fingerprinting has pointed at Meta’s Muse family, with no confirmed maker.

Problem

Agent stacks burn frontier-model budget on repetitive tool calls and sub-tasks. Crouton Alpha offers a free, long-context, multimodal stealth preview as a scalable execution layer.

Differentiation

Versus named open-weight models: anonymous hosted API with undisclosed maker. Versus earlier OpenRouter stealth listings: a Nov 3, 2026 free listing explicitly pitched for agent sub-tasks and a 1M context window.

Limits

• Anonymous provider—identity, pricing, and rate limits can change.
• Provider may retain prompts/completions; read Stealth Model Terms.
• Regional limits; EU in-region routing not currently available.
• Muse-family guesses are community inference, not official.

Who it’s for

Builders who want a cheap OpenRouter/OpenCode path for agent sub-tasks, tool calls, and multimodal document work—or who want to evaluate the stealth preview before a named release.

Get started

1. Use model id stealth/crouton-alpha via the OpenRouter page.
2. Try the free preview on OpenCode (and similar) and measure latency/tool-call quality.
3. Benchmark on your agent workloads before relying on the preview long-term.