Key takeaways
- Crouton Alpha (model ID stealth/crouton-alpha) has been listed on OpenRouter since November 3, 2026. It is free, accepts text, image, video, file and audio input, and lists a 1,048,576-token context window. OpenRouter says its developer has chosen to stay anonymous.
- In our five single-shot UI tasks, every file rendered with zero page errors and no network dependencies. The kanban board passed every functional check: add, drag to another column, persist across reload and delete.
- Visual quality was high, but not flawless: the dashboard’s bar chart printed unrounded axis values and a legend that does not match its bars, the mobile screen overlaps a label with a ring, and the landing page typesets prices with an oversized dollar sign.
- It was slow in our run: 66 to 350 seconds per answer, roughly 13 to 52 output tokens per second, with five requests sent at once on the free tier.
- Prompts and completions may be retained by the anonymous provider, and the model can be withdrawn without notice. Use it for throwaway prototypes and non-sensitive sub-tasks, not for customer data or a production dependency.
A free, anonymous 1M-context model appeared on November 3
OpenRouter added Crouton Alpha to its stealth collection on November 3, 2026; its registry timestamp is 01:38:50 UTC. The listing calls it a lightweight model for scalable coding, agentic and multimodal workflows, aimed at agentic sub-tasks, tool use and document analysis in high-volume, repeated work, and suggests pairing it with frontier reasoning models that plan while Crouton executes the bounded steps.
Input and output tokens cost $0. The registry lists text, image, video, file and audio input with text output, a 1,048,576-token context, a 943,718-token completion ceiling, tool calling with automatic tool choice only, structured outputs and mandatory reasoning whose default effort is max. The single endpoint is served by a provider named only “Stealth”, with unknown quantization. OpenRouter says it routes the requests and is not the developer; AccessAllGPT does not know who built the model and does not speculate here.
Landing page: complete and polished on the first try
The first prompt asked for a landing page for “Tidewater”, a fictional booking app for independent bike repair shops, with a hero and product mockup, three features, three-plan pricing, quotes and a footer. Crouton Alpha returned 28 KB of HTML in 142 seconds. The 4,232-pixel page contains every requested section plus a how-it-works band and a closing call to action, with a believable in-app schedule built from HTML and CSS rather than an image.
The visible flaw is typographic: each price renders as a very large dollar sign followed by a much smaller number. Like most generated marketing pages, it also invents social proof, such as “2,400+ shops” and named customer quotes, which a real team would have to replace.
Dashboard: strong layout, two chart bugs
The dashboard prompt asked for a coffee-bean store’s operations view with a sidebar, four KPI tiles, a 30-day revenue line chart and an orders-by-channel bar chart drawn in SVG, and a recent-orders table. The result, 21 KB after 240 seconds, has a coherent warm palette, consistent sample data and status badges that read well.
Two details are wrong. The bar chart’s y-axis prints raw values such as 373.33333 and 186.66666, which run into the first bar. Its legend names three channels, Website, Social and Wholesale, while the bars show four: Website, Instagram, Wholesale and Marketplace. Neither bug breaks the page, but both are the kind a reviewer must catch.
Mobile screen and kanban board: both did what was asked
The fitness-app home screen took 109 seconds. Rendered with true 390-pixel mobile emulation, it fits the width with no horizontal overflow and includes activity rings, a weekly chart, recent runs and a tab bar. Its numbers hang together: 8.2 km/h is shown as about 7′18″ per kilometre. The one visible defect is the “KM TODAY” label, which sits under the inner ring.
The kanban board was the slowest answer at 350 seconds, but it was the most important result. In the rendered page, adding a high-priority card put it in To do, a drag to Done moved it, the move survived a full reload through localStorage, and deleting it removed it permanently. We simulated the drag with in-page HTML5 drag events, not a physical mouse.
Screenshot to code: a near copy of our own home page
The fifth task sent a screenshot of the AccessAllGPT home page and asked Crouton Alpha to rebuild it. In 66 seconds it returned 7 KB of HTML that reproduces the layout, the dark top bar, the navigation, the oversized headline, the editor’s note and the first two story cards, with the text copied accurately. The fonts differ slightly because the rebuild uses system fonts.
OpenRouter’s usage record for this request reported only 99 prompt tokens, which appears to exclude the image. The output reproduced text that appeared only in the screenshot, so the image was clearly read. Teams tracking cost or context on multimodal calls should not rely on that field for this model.
Speed, cost and the stealth-model catch
Every request cost $0. Answers took 66 to 350 seconds and contained 1,888 to 7,416 output tokens, about 13 to 52 tokens per second. We sent the five requests at once, so free-tier queueing may explain part of the delay; this was not a controlled speed benchmark. Despite the max default, OpenRouter reported only 0 to 58 reasoning tokens per answer.
The price is paid in data. OpenRouter’s stealth terms say these models are free in exchange for user content, that inputs including personal data go to the anonymous provider, and that a model can be withdrawn at any time. Crouton Alpha’s listing says prompts and completions may be retained but are not used for training. The terms also forbid submitting information whose distribution law or regulation restricts.
What builders should do next
Trial Crouton Alpha for disposable work: UI mockups, prototypes, internal demos and bounded agent sub-tasks over public or synthetic data. Its one-shot front-end output is good enough to start from, and its zero price makes side-by-side comparison with a paid model cheap.
Constrain it to inputs you would be comfortable publishing, and review charts, numbers and copy before anything ships. Wait for a named developer, a price and a stable model before building a product on it. Reject it for customer data, proprietary code, regulated information or any workflow that needs an uptime or retention commitment.
Copy-ready stealth model trial record
Use one record per workload so a free preview is judged on your own tasks and data rules.
Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].
Task type, expected output, volume, latency tolerance and what happens if the model disappears tomorrow.
Which inputs may go to an anonymous provider that can retain them; confirm no personal, customer, proprietary or regulated data.
Model ID stealth/crouton-alpha, reasoning effort, max tokens, tools, structured output and input modalities used.
Frozen prompts, reference outputs or screenshots, and a paid model run on the same prompts for comparison.
Rendering errors, interaction tests, visual defects, factual or numeric errors in generated data and copy.
Pass rate, time per answer, output tokens, retries and reviewer time to fix each result.
Fallback model, how prompts move, and what changes when the provider is named or the model becomes paid.
Trial, constrain, wait or reject; owner, data rule and date to re-check the listing.
Primary sources
Browse the publication-wide evidence index →
- Crouton Alpha model pageOpenRouter · Reviewed: Model description; stealth-model notice; listing date; context window; input and output prices · Retrieved · Supports: OpenRouter lists Crouton Alpha as a free stealth model from Nov 3, 2026 with a 1,048,576-token context, describes it as a lightweight coding, agentic and multimodal model, and says an anonymous third-party provider may retain prompts and completions but does not use them for training.
- OpenRouter Models API entry for stealth/crouton-alphaOpenRouter · Reviewed: id, created timestamp, architecture and modalities, pricing, top provider limits, supported parameters, reasoning settings, endpoints · Retrieved · Supports: The registry records creation at 01:38:50 UTC on Nov 3, 2026, text, image, video, file and audio input with text output, $0 prompt and completion prices, a 943,718-token completion ceiling, tool support and mandatory reasoning with a default effort of max.
- Crouton Alpha endpointsOpenRouter · Reviewed: Provider name, quantization, tool-choice support, implicit caching, uptime fields · Retrieved · Supports: The single endpoint is served by a provider named only “Stealth” with unknown quantization, auto-only tool choice and no implicit caching; it reported 97.6% uptime over the previous day when retrieved.
- Stealth Program End User License Agreement (updated September 14, 2026)OpenRouter · Reviewed: Program overview; availability; inputs and personal data; payment; licence to user content; acceptable use policy · Retrieved · Supports: Stealth models are free in exchange for user content, may be withdrawn at any time, pass personal data in inputs to the anonymous provider, and must not receive information that law or regulation restricts.
Limitations
This is one small hands-on test, not a benchmark. Each of the five prompts was run once with default settings, so it says nothing about variance, and a second run could be better or worse. We tested only self-contained HTML front-end tasks; we did not test tool calling, structured output, long-context retrieval, video, audio or file input, agent loops, back-end code or security. Interaction checks covered only the kanban board and used synthetic drag events. Timings come from five simultaneous requests on the free tier at one point in time and are not a speed measurement. We did not compare Crouton Alpha with another model on the same prompts. The developer is anonymous, the model may change or be withdrawn, and the listing and terms may change after November 5, 2026.
Disclosures
AccessAllGPT used a standard OpenRouter account and the model’s free tier; it received no credits, early access, briefing or payment, and neither OpenRouter nor the anonymous provider sponsored, reviewed or endorsed this article. The prompts and the screenshot we sent contained no personal or confidential data. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenRouter. Publication-wide relationships are listed on the disclosures page.
