The OxAlpha FAQ — every OxAlpha question, answered

About OxAlpha

Start here if OxAlpha is new to you. These answers cover what the model is, where it came from, and why a frontier-grade system is sitting on OpenRouter with no name attached.

What is OxAlpha?

OxAlpha is a stealth AI reasoning model that appeared on OpenRouter with no announcement, no named creator, and no price tag. OxAlpha reads up to one million tokens of context, produces up to 131K output tokens, and reasons internally before it writes an answer. The community treats OxAlpha as a frontier-class model because of its results on shared reasoning suites. This site collects everything known about OxAlpha in one place and lets you chat with the model free.

Who created OxAlpha?

Nobody knows, and that is the honest answer. OxAlpha surfaced quietly under the model id stealth/ox-alpha with no lab attribution, no paper, and no press release. Stealth releases like OxAlpha are a common way for labs to gather real-world feedback before a public launch, so the leading theory is that OxAlpha is a pre-release frontier model being tested anonymously. Until someone claims it, the origin of OxAlpha remains an open question the community keeps digging into.

Why is OxAlpha called a stealth model?

A stealth model is one released without identity: no launch event, no branding, no documentation from its maker. OxAlpha fits that pattern exactly. It appeared on OpenRouter one day, free to call, with capabilities that clearly exceed a hobby project. Everything the community knows about OxAlpha comes from testing it directly, not from official sources. That is also why every OxAlpha benchmark number you see is a community measurement rather than a vendor claim.

What does thinking-first mean for OxAlpha?

OxAlpha produces internal reasoning before it writes its visible answer. Instead of predicting a reply token by token straight away, OxAlpha first works through the problem in a hidden reasoning phase, checking steps and exploring alternatives, then commits to a final response. This thinking-first design is why OxAlpha handles multi-step logic, long proofs, and tricky edge cases better than models that answer immediately. The trade-off is a short pause before OxAlpha starts responding.

Is this site or OxAlpha affiliated with OpenRouter or an AI lab?

No. This site is an independent community resource about OxAlpha. It is not run by OpenRouter, not run by whoever trained OxAlpha, and not affiliated with any AI lab. Everything published here comes from public information and hands-on community testing of OxAlpha. If the model is ever officially claimed by a lab, this site will report that, but it has no inside knowledge. You can read more about the project on the about page.

Using OxAlpha free

OxAlpha costs nothing to use, which raises its own set of questions. Here is how free access works, what you need to get started, where the practical limits are, and what the likely trade-offs behind a free frontier model look like.

How do I try OxAlpha for free?

Two ways, both free. The fastest is the chat on this site: open the OxAlpha chat and start talking to OxAlpha immediately, with no sign-up, no waitlist, and no payment details. The second way is OpenRouter, where OxAlpha is listed as stealth/ox-alpha at $0 per token. The site chat is best for quick, hands-on testing, while OpenRouter suits developers who want OxAlpha inside their own applications and scripts.

Do I need an account to chat with OxAlpha?

No account is needed to use the OxAlpha chat on this site. There is no sign-up form, no email verification, and no credit card step. You open the page and OxAlpha is ready. If you call OxAlpha through the OpenRouter API instead, you need a free OpenRouter account to get an API key, but the OxAlpha model itself still costs nothing to use through that route either.

Is there a catch to OxAlpha being free?

The realistic catch is that stealth models are usually temporary. Whoever runs OxAlpha is most likely paying for compute in exchange for real-world usage data during a testing window, and that window can close without warning. OxAlpha could be renamed, put behind a price, or withdrawn at any time. There is no service guarantee and no support channel. Enjoy OxAlpha while it lasts, and avoid building anything critical on the assumption it stays free forever.

What can I actually do with OxAlpha in the free chat?

Anything you would ask a frontier model. People use the free OxAlpha chat to debug large codebases, draft and refactor long documents, work through math and logic problems, plan multi-step projects, and analyze pasted data. Because OxAlpha accepts up to a million tokens of context, you can paste entire repositories or book-length texts into a single OxAlpha conversation and ask questions that span all of it, which most free chat tools simply cannot do.

OxAlpha API access

Developers can call OxAlpha programmatically today. These answers cover the model id, the endpoint, and the capabilities OxAlpha exposes through the API. If you already use an OpenAI-compatible client library, setup usually takes less than five minutes and one changed configuration line.

How do I call OxAlpha through the OpenRouter API?

Create a free OpenRouter account, generate an API key, and send standard chat-completion requests with the model field set to stealth/ox-alpha. OpenRouter exposes OxAlpha through an OpenAI-compatible endpoint, so most existing SDKs and libraries work with OxAlpha after a one-line change to the model name and base URL. Requests to OxAlpha are billed at $0, so the only practical constraints are rate limits imposed by the router, not your wallet.

What is the OxAlpha model id?

The model id is stealth/ox-alpha. The stealth prefix is the namespace OpenRouter uses for anonymous models, and ox-alpha is the only public identifier its creators have given the world. If you see references to OxAlpha in community posts, benchmark threads, or tooling configs, they all resolve to this one id. There are currently no size variants, no mini edition, and no dated snapshots of OxAlpha.

Does OxAlpha support tool calling and structured output?

Yes. OxAlpha handles function and tool calling reliably, choosing sensible tools and formatting arguments correctly across long multi-step chains. OxAlpha also produces clean structured output: ask for JSON matching a schema and it conforms with unusually few formatting errors. Combined with the reasoning phase, this makes OxAlpha a strong engine for agent loops, extraction pipelines, and automation where a malformed response would break downstream code.

What are the context and output limits of OxAlpha?

OxAlpha accepts up to 1,000,000 tokens of input context and can generate up to 131,000 output tokens in a single response. In practice that means OxAlpha can hold a large monorepo, a legal archive, or several novels in working memory at once, and can answer with an entire report or a large code patch in one pass. Few models at any price match this combination, and OxAlpha offers it at no cost.

OxAlpha performance

Numbers first, caveats attached: every figure below is a community-run measurement, because OxAlpha has no official documentation to cite. Where possible, the linked pages point to the raw testing threads, so you can inspect the prompts, the harnesses, and the grading criteria yourself before trusting any single number.

How does OxAlpha score on benchmarks?

On an aggregate community reasoning suite, OxAlpha scores roughly 80%, ahead of GPT-5 at about 65%, Fable 5 at about 62%, and Grok 4 at about 58%. Treat these as directional community measurements, not official scores: prompts, harnesses, and scoring choices all move results by a few points. Even with that caveat, OxAlpha leads by a margin large enough that methodology alone cannot explain it. Full details are on the OxAlpha benchmarks page.

How does OxAlpha compare with GPT-5 and Fable 5?

In community testing, OxAlpha wins most head-to-head reasoning and long-context tasks, while the named models remain competitive on speed and ecosystem maturity. GPT-5 and Fable 5 ship with official support, stable availability, and documented behavior, none of which an anonymous stealth release can promise. See the detailed OxAlpha vs GPT-5 comparison for task-by-task results. The short version: OxAlpha is the stronger reasoner today, the incumbents are the safer dependency.

Is OxAlpha good at coding?

Long-horizon coding is one of the clearest strengths of OxAlpha. Testers report that OxAlpha stays coherent across large multi-file changes, remembers architectural decisions made hundreds of thousands of tokens earlier, and debugs by actually reasoning through execution paths instead of pattern-matching error strings. The 131K output ceiling means the model can emit a complete refactor, tests included, in one response instead of a truncated fragment. For repository-scale work, the huge context makes OxAlpha genuinely different, not just incrementally better.

Can OxAlpha process images and video?

Yes. OxAlpha is multimodal on the input side: it accepts text, images, and video, and reasons across them together. You can hand OxAlpha a screenshot of a stack trace with the relevant source code, or a video walkthrough of a bug, and get an answer grounded in both. Output is text only. The model does not generate images or video, so pair it with a dedicated generator when you need media out of the pipeline.

Privacy and safety with OxAlpha

An anonymous model deserves extra caution. These final answers cover what happens to your data and what limits apply when you use OxAlpha. The guiding rule is simple: enjoy the capability, but share only what you would be comfortable posting publicly.

Is my data safe when I use OxAlpha?

Assume it is not private. Because the operator of OxAlpha is unknown, there is no privacy policy, no data-processing agreement, and no retention promise you can rely on. The most likely purpose of a free stealth release is collecting real usage data, so it is sensible to assume prompts sent to OxAlpha may be stored and used for training. Treat every OxAlpha conversation as if it could be read by the model’s creators.

Should I send confidential or personal data to OxAlpha?

No. Keep secrets, credentials, unreleased source code, medical records, and personal identifiers out of OxAlpha prompts. This is good practice with any hosted model, but it matters doubly with OxAlpha because there is no legal entity to hold accountable if data is mishandled. For confidential workloads, use a provider with a signed agreement. For everything else, public code, homework, drafts, and general questions, OxAlpha is a free and capable choice.

Are there rate limits or usage caps on OxAlpha?

There is no published limit, but practical throttles exist. The free chat on this site applies light rate limiting to keep OxAlpha responsive for everyone, and OpenRouter applies its own per-key limits to free models. Heavy automated traffic may see slowdowns or temporary refusals. Since OxAlpha costs nothing, the honest advice is to be a polite tenant: batch your requests sensibly and expect occasional capacity hiccups for as long as the stealth deployment lasts.