Stealth Ox Alpha Review: Is the Free Mystery AI Model Worth It?
A hands-on test of Stealth Ox Alpha, a free anonymous AI model on OpenRouter, run inside Claude Code for coding and knowledge tasks.
What is Stealth Ox Alpha?
Stealth Ox Alpha is a free AI model distributed through OpenRouter under an anonymous provider called “Stealth.” It costs $0 per million input tokens and $0 per million output tokens, and its origin isn’t disclosed: no company name, no confirmed training data, no country of origin. Models like this show up periodically on OpenRouter as “stealth” or “cloaked” releases, often used to gather real-world usage data before a public launch. Because the provider is anonymous, anyone routing sensitive business data through it is doing so on trust alone.
The interesting part isn’t the model card, it’s that people are running it inside Claude Code, Anthropic’s coding agent harness, by swapping out the backend model. That combination, a free anonymous model wearing Claude Code’s tool-use and orchestration layer, is what’s actually being tested here.
TL;DR
- Stealth Ox Alpha is a free, anonymous model on OpenRouter that can be wired into Claude Code by editing the harness’s settings file and pointing it at an OpenRouter API key instead of Anthropic’s.
- Swapping models only works in IDE or terminal setups (VS Code plus the Claude Code extension, or a terminal session), not in the Claude Code desktop app, which hardcodes Anthropic’s models.
- In hands-on testing, the model handled knowledge work tasks reasonably well: it built a working front-end landing page, pulled a YouTube analytics report, and pulled data from a Google Sheet using stored skills and memory files.
- The biggest problem was speed, not quality: tasks that normally take Claude 5 to 30 minutes took the free model as long as 6 hours, with repeated “upstream idle timeout exceeded” errors along the way.
- The model showed some agentic resourcefulness, digging through unrelated project folders to find API keys and even scheduling a cron job to retry a blocked task on its own.
- OpenRouter’s free model lineup changes frequently, so a specific free model like Ox Alpha may disappear or stop being free at any time; OpenRouter also offers a “free models router” that auto-selects whatever free model is currently available.
- The general takeaway is that free models are viable for lighter, less time-sensitive knowledge work, but not yet a real substitute for paid frontier models like Claude Opus or GPT-5 class models on complex orchestration or software tasks.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How do you run a free OpenRouter model inside Claude Code?
The setup requires an OpenRouter account, which gives access to one API key that routes to more than 500 models, including free ones. From there:
- Create an OpenRouter account and generate an API key from the account’s “API Keys” section.
- Open Claude Code’s settings file (this only works in VS Code or a terminal, not the desktop app).
- Inside the
envsection of the settings file, replace the Anthropic API token field with the OpenRouter API key. - Set the model fields to the exact model slug shown on OpenRouter, for example
stealth/ox-alpha. - Save the file and start a new Claude Code session. The active model name should appear in the session header or status line as confirmation.
Switching models later just means swapping the model string for a different one, such as a GLM free-tier model, in every relevant slot in the settings file.
Because free listings rotate, OpenRouter also provides a generic “free models router” option that automatically selects from whatever free models are currently available, which is a more durable way to keep using $0 inference without tracking which specific model is free this week.
How fast and reliable is Stealth Ox Alpha in practice?
Speed was the model’s clearest weakness. Three separate agentic tasks, building a landing page, generating a YouTube analytics report, and processing a video into a skills database, each took roughly 6 hours or, in one case, about 45 minutes for a task that should have taken 5 minutes with standard Claude. Several runs hit “upstream idle timeout exceeded” errors, which appeared to stem from the model attempting to process or write too much at once. Breaking tasks into smaller chunks reduced but didn’t eliminate the slowdown.
Reliability was mixed rather than broken: the model didn’t produce garbage output, it just took an unpredictable amount of time to get there, and occasionally stalled entirely, requiring manual intervention or a retry loop. In one case, the model set up its own cron job to periodically retry a blocked step, which is a small but notable sign of agentic behavior baked into the harness rather than the model itself, since Claude Code’s tool access enables that regardless of which backend model is driving it.
Is Stealth Ox Alpha good enough for real work?
For coding and technical software tasks, the answer leans no. Free and smaller models running inside Claude Code are not matching what Claude Opus or GPT-5 class models can do on deep technical work, especially tasks requiring an orchestrator to manage multiple sub-agents or complex multi-step scoping.
For knowledge work, the results were more encouraging. The model:
- Built a functional front-end landing page with a working shopping cart, correct brand colors, logo usage, and product data pulled from existing files, including nutrition facts it generated on its own.
- Investigated an unrelated project folder to find previously stored API keys and scripts, then used them to pull real YouTube channel analytics without being told exactly where to look.
- Produced a quarterly and annual report covering per-video and per-month metrics, audience geography, device data, and a forward-looking content prediction, delivered as a formatted Google Sheet.
- Recognized and used existing memory files and stored skills, showing it could operate inside Claude Code’s context system rather than ignoring it.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
The quality of these outputs was described as “not bad” and comparable in structure to what Claude would produce, just far slower to get there. For simpler, lower-stakes tasks where turnaround time doesn’t matter, a free model handled the work adequately. For anything time-sensitive or technically demanding, the wait made it impractical.
Should you use an anonymous stealth model for sensitive data?
This is the part worth taking seriously before treating free stealth models as a no-downside swap. Because the provider behind Stealth Ox Alpha isn’t disclosed, there’s no way to verify data handling practices, retention policies, or where the model was trained. Anyone routing proprietary business data, customer information, or confidential code through an anonymous model is accepting that risk without the transparency a named provider like Anthropic or OpenAI offers. For testing, experimentation, or non-sensitive content generation, that tradeoff may be acceptable. For production use involving real business or customer data, it’s a meaningfully different risk profile than using a known, accountable model provider.
Frequently Asked Questions
What is Stealth Ox Alpha?
It’s a free, anonymously provided AI model available through OpenRouter. Its developer, training data, and origin aren’t publicly disclosed, and it’s listed at $0 cost per million input and output tokens.
Can you use Stealth Ox Alpha with Claude Code?
Yes, by editing Claude Code’s settings file to route through an OpenRouter API key instead of Anthropic’s, and setting the model field to the OpenRouter model slug. This only works in the VS Code extension or terminal version of Claude Code, not the desktop app.
Is Stealth Ox Alpha actually free?
It was free to use through OpenRouter at the time of testing, though OpenRouter’s free model roster changes over time and specific models can be removed or repriced without notice.
How does Stealth Ox Alpha compare to Claude on speed?
It was substantially slower. Tasks that take standard Claude models minutes took Stealth Ox Alpha hours in testing, and it produced timeout errors on larger or more complex requests.
Is Stealth Ox Alpha good for coding tasks?
Not particularly. It performed reasonably on lighter knowledge work like reports and simple front-end pages, but it wasn’t considered a strong substitute for paid frontier models like Claude Opus or GPT-5 on complex coding or orchestration work.