OxAlpha — the reasoning model that appeared out of nowhere

What is OxAlpha?

OxAlpha is a stealth reasoning model that surfaced quietly on OpenRouter under the model id stealth/ox-alpha, with no announcement and no named creator. Nobody outside its makers knows which lab trained OxAlpha, what hardware runs it, or why it is free. What the community does know is what OxAlpha can do, because thousands of developers have been testing OxAlpha since the day it appeared — and the results keep turning heads.

On paper, the OxAlpha specification is unusual for an unbranded model. OxAlpha accepts up to one million tokens of context, which is enough for an entire codebase, a stack of legal contracts, or a full book series in a single prompt. OxAlpha can produce up to 131K output tokens in one response, so long documents and large refactors come back complete instead of truncated. And OxAlpha costs nothing: the OpenRouter listing is free, and the chat on this site is free with no account.

Stealth models are not new. Labs sometimes ship anonymous checkpoints to gather unbiased feedback before a launch. What makes OxAlpha different is the gap between its anonymity and its ability. Most stealth releases feel like mid-tier previews. OxAlpha behaves like a frontier system, and in community testing OxAlpha often outperforms the flagship models it is presumably hiding among. This site is an independent community resource that documents OxAlpha — we are not affiliated with any AI lab or with OpenRouter, and you can read more on the about page.

The mystery has become part of the appeal, but it is worth being clear about what it means in practice. Nothing about OxAlpha is documented by a vendor: no model card, no training-data disclosure, no deprecation policy. Everything this site reports about OxAlpha comes from direct observation — running prompts, measuring outputs, and comparing notes across the community. That makes the picture of OxAlpha unusually honest in one sense (nobody is marketing it to you) and unusually fragile in another (OxAlpha could be renamed, gated, or withdrawn tomorrow). We update this site as the facts change, and we label every number with where it came from.

OxAlpha reasoning: it thinks before it speaks

The defining trait of OxAlpha is that it is thinking-first. Before OxAlpha writes a single word of its answer, it produces an internal chain of reasoning: it decomposes the problem, weighs candidate approaches, checks its own intermediate steps, and only then commits to a response. You can watch this happen in the chat — OxAlpha visibly pauses to reason, then answers with unusual precision.

This design changes the failure modes people expect from language models. Where a fast autocomplete-style model guesses and patches, OxAlpha plans. Ask OxAlpha a multi-step logic puzzle and it works the steps in order rather than pattern-matching to a similar puzzle. Give OxAlpha a subtle bug across three files and it traces the actual data flow before proposing a fix. The reasoning phase costs a few seconds of latency, but it is precisely why OxAlpha scores so far ahead on reasoning-heavy evaluations.

The long context window compounds this strength. Because OxAlpha can hold a million tokens in view while it reasons, it does not lose the thread on long-horizon tasks: hour-long agent runs, multi-file refactors, or analysis that references material from hundreds of pages back. If you want the deeper technical picture of how the reasoning loop works, see how OxAlpha works.

There is a useful mental model here: treat OxAlpha less like a chatbot and more like a careful colleague. A chatbot rewards short, conversational prompts. OxAlpha rewards well-specified problems — paste the whole ticket, the whole file, the whole constraint list — because the reasoning phase gets stronger as the input gets richer. Users who bring OxAlpha real context report dramatically better results than users who feed it one-line questions, and that difference shows up in every category the community measures.

OxAlpha capabilities at a glance

OxAlpha is a generalist, but community testing shows four areas where OxAlpha clearly separates itself from the pack. The capabilities page covers each in depth; here is the short version. A pattern worth noticing: none of these are party tricks. Each strength of OxAlpha maps to a real workflow — shipping software, auditing documents, running agents — which is why OxAlpha adoption has spread through working engineers rather than through hype.

Long-horizon coding

OxAlpha sustains coherent work across very large codebases and very long sessions. Developers report that OxAlpha completes multi-hour agentic coding tasks — plan, edit, test, repeat — without drifting from the original goal, and that its patches respect surrounding architecture instead of fighting it. Combined with the 1M-token window, OxAlpha can hold an entire monorepo in context while it edits, which removes a whole class of wrong-because-it-could-not-see-it mistakes.

Complex reasoning

Mathematics, formal logic, scientific analysis, and adversarial word problems are where the thinking-first design of OxAlpha pays off most. OxAlpha shows its work, catches its own mistakes mid-derivation, and holds up on problems that make other models hallucinate a confident wrong answer.

Tool calling and structured output

OxAlpha emits clean, schema-conforming JSON and chains tool calls reliably, which makes OxAlpha a strong backbone for agents and pipelines. Function arguments arrive valid, and OxAlpha recovers gracefully when a tool returns an error.

Multimodal input

OxAlpha accepts text, images, and video as input. Screenshots of stack traces, whiteboard photos, UI mockups, and recorded screen captures can all go straight into the prompt, and OxAlpha reasons over them together with the surrounding text. In practice this means you can hand OxAlpha a screen recording of a bug plus the relevant source files and ask for a diagnosis in one shot — a workflow that used to require three separate tools.

OxAlpha benchmarks: the short version

Because no lab has claimed OxAlpha, there are no official scores — every number you will see is community-run and should be read as directional rather than definitive. With that caveat stated plainly, the direction is hard to ignore. On an aggregate community reasoning suite, OxAlpha lands around 80%, well clear of the named frontier models tested under the same conditions.

A fifteen-point lead over GPT-5 on reasoning-heavy tasks is the kind of gap that usually marks a generation change, not a point release. For the full methodology, per-category breakdowns, and honest notes on where OxAlpha does not win, see the dedicated OxAlpha benchmarks page, or the head-to-head comparisons of OxAlpha vs GPT-5 and OxAlpha vs Fable 5.

Two caveats keep these numbers honest. First, community suites skew toward exactly the reasoning-heavy tasks where the thinking-first design of OxAlpha shines; a benchmark weighted toward speed or creative writing would narrow the gap. Second, sample sizes vary by category, so treat single-category results as noisier than the aggregate. Even with both caveats applied, OxAlpha has now topped every major community leaderboard it has appeared on, across independent testers who do not share methodology — and that consistency, more than any single score, is what convinced skeptics that OxAlpha is a genuine frontier model rather than a fine-tune with good marketing.

How to use OxAlpha free

There are two ways to run OxAlpha today, and both cost nothing. That sentence sounds like a catch, so let us be precise: there is no trial period, no token allowance that runs out, and no premium tier hiding behind the free one. Whoever operates OxAlpha is absorbing the compute bill, presumably in exchange for the real-world usage data that stealth releases exist to collect.

Option one: chat here. The free OxAlpha chat on this site talks to the live model with no sign-up, no email, and no credit card. Open the page, type a prompt, and OxAlpha responds — reasoning phase and all. It is the fastest way to form your own opinion of OxAlpha, and it takes under a minute.

Option two: the API via OpenRouter. Developers can call OxAlpha through OpenRouter using the model id stealth/ox-alpha. The listing is free, it speaks the standard OpenAI-compatible API shape, and it supports streaming, tool calling, and structured output. Point your existing client at the OxAlpha model id and your agents, scripts, and evaluations run unchanged.

One practical warning applies to any stealth model: assume prompts may be logged by whoever operates OxAlpha, and keep genuinely sensitive data out. For setup details, rate-limit notes, and prompt tips that play to the strengths of OxAlpha, read how OxAlpha works.

OxAlpha FAQ

Who made OxAlpha?

Nobody knows, and no lab has claimed it. OxAlpha appeared on OpenRouter as an anonymous stealth listing, and its behavior does not cleanly fingerprint any one vendor. The most common theory is that OxAlpha is a pre-release frontier checkpoint gathering real-world feedback, but that remains speculation. Guessing games aside, the anonymity cuts both ways: OxAlpha earns its reputation purely on output, with no brand loyalty inflating opinions in either direction.

Is OxAlpha really free?

Yes. The OpenRouter listing for OxAlpha is priced at $0 for both input and output tokens, and the chat on this site is free with no account. Stealth listings can change or disappear without notice, so the honest framing is: OxAlpha is free today, with no guarantee about tomorrow.

How large is the OxAlpha context window?

OxAlpha accepts up to one million tokens of input context and can generate up to 131K output tokens in a single response. That puts OxAlpha at the top of its class on both numbers, and it is a big part of why long-horizon coding is such a strength. For scale: a million tokens is roughly 750,000 words, so OxAlpha can take an entire large repository or several novels of material in a single request.

Is OxAlpha better than GPT-5?

On community reasoning suites, OxAlpha scores around 80% against roughly 65% for GPT-5 — a large, consistent lead. These are unofficial measurements, and GPT-5 still competes closely in some categories, so read the full comparison before declaring a winner for your use case.

Can OxAlpha see images and video?

Yes. OxAlpha is multimodal on the input side: it accepts text, images, and video, and it reasons across all of them together. Output is text only — OxAlpha does not generate images or video. Community testing shows OxAlpha is particularly good at reading dense technical images such as diagrams, charts, and screenshots of code.

Is it safe to use OxAlpha for real work?

For code, analysis, and drafting, many teams already do. Because the operator of OxAlpha is unknown, treat it like any third-party service: keep secrets and regulated data out of prompts. A sensible policy is simple: use OxAlpha freely for anything you would post publicly or share with a contractor, and hold back anything you would not. More questions about OxAlpha, its limits, and its quirks are answered on the full FAQ page.