OxAlpha capabilities: what the stealth model can actually do

A note on sourcing before the tour begins. Nobody has published an official OxAlpha model card, so everything below comes from community testing: independent users have probed OxAlpha capability by capability and compared notes in public. Where a number appears, it is a community measurement. Where a use case appears, someone has actually pushed it through OxAlpha. The scores behind these claims live on the OxAlpha benchmarks page; this page stays focused on what OxAlpha does and where each OxAlpha strength pays off in day-to-day work.

Long-horizon coding with OxAlpha

The capability that put OxAlpha on the map is sustained, goal-directed coding. Plenty of models can write a function. OxAlpha can carry a plan across hours of agentic work — reading a ticket, exploring a codebase, editing files, running tests, reading the failures, and iterating — without forgetting what the original goal was. Community testers describe the difference bluntly: other models drift, OxAlpha does not.

Whole-repository refactors

Because OxAlpha accepts up to a million tokens of context, you can hand OxAlpha an entire mid-sized repository and ask for a change that touches dozens of files: rename a core abstraction, migrate an API client, or unwind a circular dependency. With the whole codebase in view, OxAlpha keeps call sites, tests, and type definitions consistent, which removes the classic failure where a model fixes three files and silently breaks seven others it never saw.

Multi-hour agent sessions

In agent harnesses, OxAlpha shows unusual patience. OxAlpha plans before editing, verifies after editing, and treats a failing test as information rather than as a reason to thrash. One popular community stress test gives models a vague feature request and a large open-source project; OxAlpha routinely produces a merged- quality pull request where competing models stall or ship regressions. The benchmarks page breaks the OxAlpha coding results down category by category.

Concrete use cases where OxAlpha earns its keep: dependency upgrades that ripple through a codebase, test-suite backfills for legacy modules, framework migrations, and bug hunts where the symptom and the cause live in different services. In each case the pattern is the same — the task is long, the context is large, and OxAlpha holds both without dropping either. Teams also use OxAlpha as a reviewer: because OxAlpha reads the entire diff and the entire codebase together, an OxAlpha review catches cross-file inconsistencies that a per-file reviewer never sees.

Complex reasoning: the thinking-first core of OxAlpha

Every other OxAlpha capability sits on top of one design choice: OxAlpha reasons before it responds. Ask OxAlpha a question and OxAlpha first produces an internal chain of thought — decomposing the problem, weighing approaches, checking intermediate steps — and only then writes the answer. The pause is visible when you use OxAlpha in the free chat, and the payoff is visible in the output.

On community reasoning suites, this discipline translates into an aggregate score around 80% for OxAlpha, against roughly 65% for GPT-5 and 62% for Fable 5. Those are unofficial, community-run measurements, not vendor scores, but the gap is consistent across independent testers. Where the thinking-first approach helps most:

• Mathematics and formal logic. OxAlpha works derivations step by step and catches its own arithmetic slips mid-stream instead of presenting a confident wrong answer.
• Adversarial word problems. Trick questions designed to trigger pattern-matching tend to fail against OxAlpha, because OxAlpha re-derives the problem rather than recognizing it.
• Analysis under constraints. Give OxAlpha ten interacting requirements and OxAlpha tracks all ten, flagging conflicts between them instead of quietly satisfying seven.

One underrated OxAlpha behavior deserves its own mention: when the evidence is thin, OxAlpha hedges explicitly instead of inventing certainty. That honesty makes OxAlpha output easier to trust in high-stakes analysis, where a confident fabrication is worse than an admitted gap.

The practical implication is that OxAlpha rewards rich prompts. The more constraints, context, and raw material you provide, the more the reasoning phase has to work with. A one-line question gets a good answer; a full specification gets a remarkable one. The mechanics of the reasoning loop are covered in depth in how OxAlpha works.

The 1M-token context window

OxAlpha reads up to one million tokens of input — roughly 750,000 words — and produces up to 131K tokens of output in a single response. Both numbers matter. The input side means OxAlpha can take an entire codebase, a year of meeting notes, or a shelf of contracts in one prompt. The output side means the answer comes back whole: ask OxAlpha for a complete migration guide or a full test suite and the 131K output budget lets OxAlpha finish the job in one response, where other models stop at a fraction of the length and ask permission to continue.

What actually fits in a million tokens

For intuition: a million tokens is about 4,000 pages of text, a typical 50,000-file monorepo's core source, or several novels. Community users routinely load OxAlpha with inputs that would require chunking, embedding, and retrieval pipelines on other models, and simply ask their question directly. That is not just convenient — it is more accurate, because OxAlpha sees the real document instead of a retrieval system's guess about which fragments matter.

Long-context recall in practice

A big window is only useful if the model can find things in it. Needle-in-a-haystack tests run by the community show OxAlpha retrieving specific facts from deep positions in near-full contexts with high reliability, and — harder — synthesizing claims scattered across hundreds of pages into one coherent answer. Typical use cases: cross-referencing a legal contract against a regulation, auditing a codebase for every place a deprecated pattern appears, or summarizing a research corpus without losing the outlier papers.

Tool calling and structured output

OxAlpha is unusually dependable as the engine inside automated systems, which is where sloppy models get expensive. Two related capabilities drive this: disciplined tool calling and clean structured output.

Tool calling for agents

OxAlpha emits well-formed function calls with valid arguments, chains them across long sequences, and — critically — recovers when a tool returns an error instead of hallucinating a success. In multi-tool agent loops (search, then read, then compute, then write), community testers report that OxAlpha picks the right tool for the step at hand rather than hammering one favorite. That makes OxAlpha a strong backbone for research agents, customer-support automations, and data pipelines that call real APIs with real consequences. A useful detail for agent builders: OxAlpha keeps its reasoning phase active between tool calls, so OxAlpha decides the next step from the actual tool result rather than from a script it committed to at the start.

JSON mode and schema conformance

Ask OxAlpha for JSON matching a schema and you get JSON matching the schema — first try, no markdown fences, no trailing commentary. For anyone building extraction pipelines this is the difference between a working product and a retry loop. Use cases the community has documented: converting unstructured documents into database rows, extracting entities and relationships from reports, normalizing scraped data, and generating configuration files that validate on the first pass. Because OxAlpha is free through OpenRouter under the id stealth/ox-alpha, teams run these high-volume extraction jobs through OxAlpha at zero marginal cost.

Multimodal input and production workloads

OxAlpha accepts text, images, and video in the same prompt and reasons across all of them together. Output is text only — OxAlpha does not generate images — but on the input side the flexibility changes real workflows, because visual evidence stops being a separate pipeline and becomes ordinary context.

Images and video as first-class context

Screenshots of stack traces, photos of whiteboard architecture sketches, UI mockups, dense charts, and recorded screen captures can all go straight into an OxAlpha prompt. The community consensus is that OxAlpha is particularly strong on technical images — OxAlpha reads code in screenshots accurately and follows lines in diagrams correctly. A representative workflow: attach a screen recording of a bug, paste the relevant source files beside it, and ask OxAlpha to connect the visible symptom to the offending line. That diagnosis used to take three tools and a human in the middle; OxAlpha does it in one shot.

Is OxAlpha ready for production?

Capability-wise, yes — teams already run OxAlpha inside coding agents, extraction pipelines, and internal analysis tools. Operationally, the honest answer is more careful. OxAlpha is a stealth model with no SLA, no deprecation policy, and an unknown operator, so sensible teams treat OxAlpha the way they would treat any free third-party service: keep secrets and regulated data out of prompts, wrap OxAlpha calls with a fallback model, and assume the listing could change without notice. Within those guardrails, the economics are hard to argue with — OxAlpha delivers frontier-class output at $0. This site is an independent community resource, unaffiliated with any lab or with OpenRouter, and every OxAlpha claim here is labeled with where it came from.

OxAlpha capabilities FAQ

What is OxAlpha best at?

Long-horizon coding and complex reasoning are the standout OxAlpha capabilities. On community evaluations, the gap over named frontier models is widest on tasks that demand sustained, multi-step thinking — large refactors, mathematical derivations, and constraint-heavy analysis. Tool calling and structured output are close behind, which is why OxAlpha shows up so often inside agent frameworks. If you only test one thing, test OxAlpha on a refactor you already know well.

How long can OxAlpha work on a coding task?

Community agent runs show OxAlpha sustaining coherent multi-hour sessions — planning, editing, testing, and iterating — without losing the original objective. The 1M-token window helps: the full history of the session and the full codebase both stay in view, so OxAlpha never has to act on a summary of its own past work.

Does OxAlpha support JSON mode and function calling?

Yes. OxAlpha produces schema-conforming JSON reliably and handles multi-step tool calling, including graceful recovery from tool errors. Through OpenRouter, OxAlpha speaks the standard OpenAI-compatible API shape, so existing function-calling code works against the stealth/ox-alpha id without modification. In practice that means you can drop OxAlpha into an existing agent stack and compare OxAlpha against your current model in an afternoon.

Can OxAlpha understand video?

Yes, as input. You can attach video — screen recordings are the most common community use — alongside text and images, and OxAlpha reasons over all of it together. OxAlpha does not generate video or images; OxAlpha output is text only.

How does the OxAlpha context window compare to other models?

At one million input tokens and 131K output tokens, OxAlpha sits at or above the top of the frontier class on both figures. More important than the raw number is recall quality: community needle-in-a-haystack testing shows OxAlpha actually using the deep end of its window rather than merely accepting it. The benchmarks page has the detailed OxAlpha long-context results.

Does using every OxAlpha capability cost anything?

No. Every capability described here — long context, tool calling, JSON mode, multimodal input — is available on the free OpenRouter listing and in the free chat on this site, with no sign-up and no tier system. Stealth listings can change without notice, so the precise claim is that OxAlpha is free today. More general questions are answered on the FAQ page.