Quick Answer
Ox Alpha (stealth/ox-alpha) is a preview-period reasoning model on OpenRouter, currently free on the listing, with a 1,048,576-token (1M) context window, up to 131,072 output tokens, and text + image + video input. A practical first task looks like this: pick one contained, non-sensitive coding or analysis problem; make reasoning effort a deliberate choice; feed a well-ordered long context; use tools and structured output only where they earn their latency; and grade the result with a small repeatable evaluation you can rerun later. You can try Ox Alpha free on glm5.app right now; this guide covers what to do after you have access.
Why “After Access” Is the Hard Part
Coverage of Ox Alpha has focused on access: preview pricing, API parameters, and the limited-time OpenCode Go offer. Less covered is what happens after the key works: choosing a safe first task, sizing reasoning effort, preparing a long context, and deciding when to use tools, structured output, or multimodal input.
That gap matters because Ox Alpha’s official listing has three unusual properties: reasoning is mandatory, the provider is anonymous, and there are no official benchmark scores. Each changes how you run a first real workload. All facts below come from the OpenRouter official model listing, the public OpenRouter API, OpenRouter’s Stealth Model Terms, and public announcements, checked on August 23, 2026.
Choose a Safe First Task
OpenRouter positions Ox Alpha as a reasoning model for coding, sustained agentic work, and production workloads, including workflows that combine text with visual context. That tells you what the model is aimed at, not which task should be first.
Early usage signals point in the same direction. The public model page shows heavy usage by agentic coding tools, including Hermes Agent, Claude Code, Oh-My-Pi, DeepSeek Harness, and ZCode. Start with a coding or multi-step reasoning task rather than trivia.
Use three filters for a first task:
- Contained scope: one repository module, one refactor, or one analysis document—not an open-ended project.
- No sensitive data: the Stealth terms state that the provider retains prompts and completions. Use synthetic, public, or redacted material first.
- Reproducible outcome: choose something with a ground-truth answer so you can judge whether the result is right.
Set Reasoning Effort Deliberately
Ox Alpha has mandatory reasoning. It cannot be switched off; the default effort is max, and the API supports max, high, and low. That makes effort choice part of task design.
- Use
maxfor the hard core of a task: multi-step deduction, long-horizon planning, or repository-wide reasoning. - Use
lowfor mechanical work such as extraction, formatting, or summarising material you already understand. - Keep the setting fixed during an evaluation. The API also exposes
temperature,top_p, andtop_k; stable parameters make repeated runs comparable.
OpenRouter’s public monitoring currently shows roughly 5.81 seconds P50 latency and 24 tokens per second P50 throughput. Long, max-effort requests can take minutes. Treat slow responses as a planning constraint, not automatically as a failure.
Structure a Long-Context Request
Ox Alpha supports 1,048,576 input tokens and up to 131,072 output tokens. That is headroom for genuinely large inputs and long artifacts—not a reason to paste everything blindly.
For a first task, order the request like this:
- Instructions first: role, constraints, expected format, and acceptance criteria.
- Material second: repository layout or document index, then relevant sections without duplicates.
- One deliverable last: make the required result explicit and testable.
A well-organised 50K-token context is usually more useful than a padded 200K-token one. For very large material, do a retrieval pass first: identify relevant files or sections, then provide only the selected content.
When to Use Tools and Structured Output
The API supports tools, tool_choice, response_format, include_reasoning, max_tokens, reasoning_effort, temperature, top_k, and top_p. Tool calling and structured output are available, but neither belongs in every request.
Use tools when correctness depends on fresh state: reading files, running tests, querying an API, or checking a schema. Use response_format when a downstream script or agent must parse the result.
Do not add them by default. Each tool round trip adds latency, and an unnecessary schema creates another failure mode. A useful first-task rule is simple: no tools when the answer is fully contained in the prompt; tools when the task must observe changing state; structured output when another system parses the result.
Use Image and Video Input Safely
The model accepts text, image, and video input and returns text. Screenshots, UI mockups, diagrams, and short clips are natural multimodal tests.
The safety boundary is the same as for text, but easier to forget: images and videos are prompts too. A dashboard screenshot can contain customer data, internal names, tokens, or proprietary UI. Start with public screenshots, synthetic mockups, or redacted captures that you have permission to share.
Respect the Anonymous-Provider Data Boundary
Ox Alpha is a stealth model. OpenRouter states that it is developed and operated by an anonymous third-party provider, that OpenRouter routes requests but is not the developer or owner, and that prompts and completions are retained by the provider but not used for training.
For a first task, that means:
- Treat the provider as a new vendor with no disclosed identity, architecture, parameter count, or post-preview price.
- Do not send secrets, credentials, customer data, or unreleased code.
- OpenRouter routes directly to a single provider called Stealth, so there is no routing ambiguity or automatic provider diversity.
- Preview pricing and terms can change; do not make an unmonitored production dependency from a first experiment.
What Most Ox Alpha Tutorials Miss
Most tutorials explain how to gain access or repeat the spec sheet. The operational details below matter more once you start using the model.
- Reasoning is a dial you cannot turn off. Mandatory reasoning with a
maxdefault means effort selection is part of prompt design, not a minor API option. - There are no official benchmarks. The model listing has no intelligence, coding, or agentic score. Any community number is community-reported and not independently verified, so your own repeatable evaluation is the dependable signal.
- Usage data reveals the real workload. Agentic coding tools dominate the public usage signals. A long-horizon, tool-using task is a more informative test than a one-shot question.
- The data boundary is part of the product. An anonymous single provider and retained prompts are constraints on what you may send, not a footnote.
- Free is a preview state. The $0/$0 listing price can change, and post-preview pricing is undisclosed. Establish a baseline while price is not yet a variable.
- Output capacity matters. A 131,072-token output ceiling can support detailed plans, full-file drafts, and multi-part artifacts. A single paragraph does not exercise the model’s intended workload.
Run a Repeatable Mini-Evaluation
Because there are no official scores, make your own evaluation small and fixed:
- Pick 3–5 representative tasks: a refactor, a long-context question, a visual-input task, and a tool-use task where appropriate.
- Fix the inputs: identical prompts, context, tools, and reasoning effort.
- Use a short rubric: correctness, instruction following, output-format compliance, and sensible tool use.
- Record outputs and retain the task set so you can rerun it after provider changes.
Run the baseline while preview pricing remains free. If you want a zero-setup place to begin, start a free Ox Alpha chat on glm5.app and use the same stealth/ox-alpha model for a contained first task.
FAQ
Is Ox Alpha actually free?
The OpenRouter listing currently shows $0 prompt and completion pricing during preview. Post-preview pricing is undisclosed.
Can I turn off reasoning?
No. The API reports mandatory reasoning; max, high, and low effort settings are supported.
How large are the context and output windows?
The public API lists 1,048,576 tokens of context and a maximum 131,072-token output.
Does it accept images and video?
Yes: text, image, and video input produce text output. Apply the same data-retention caution to visual inputs.
Does it support tool calling and JSON output?
Yes. Supported parameters include tools, tool_choice, and response_format alongside reasoning and sampling controls.
Are there official benchmark scores?
No. The OpenRouter listing has no benchmark scores. Treat circulating community scores as unverified unless independently reproduced.
Sources
- OpenRouter — Ox Alpha model page
- OpenRouter API — models
- OpenRouter — Stealth Model Terms
- OpenRouter documentation
- OpenCode docs — models
Ox Alpha’s preview is the least expensive time to learn its behaviour. Pick one contained task, fix the evaluation inputs, and run it—then use Ox Alpha free on glm5.app whenever you want the same model without building a harness first.
