What Is Ox Alpha?

Ox Alpha is the public name of an anonymous reasoning model available through OpenRouter under the model ID stealth/ox-alpha. The official Stealth provider page describes it as a model for coding, sustained agentic work, production workloads, complex reasoning, and tasks that combine text with visual context. That positioning tells developers what the provider wants the model used for. It is not, by itself, independent proof that the model performs at a particular level.

The unusual part is not the API shape. Ox Alpha works through the same OpenRouter interface developers use for named models. The unusual part is that the developer and operator have chosen not to identify themselves during the preview. OpenRouter explicitly says that it routes requests to a third-party provider and is not Ox Alpha’s developer, owner, or provider.

The timeline so far

The creation timestamp returned by OpenRouter’s public models API corresponds to August 20, 2026 UTC. On August 23, Bloomberg coverage syndicated by The Straits Times described the model as gaining traction among developers while its creator remained a mystery. That report is useful evidence about the public reaction. It does not upgrade any identity theory into a confirmed attribution.

What OpenRouter confirms

As of the date on this page, the public model record supports the following statements:

• The context window is 1,048,576 tokens.
• The maximum completion length exposed for the top provider is 131,072 tokens.
• Inputs can include text, images, and video; the output modality is text.
• The supported parameter list includes tools, tool choice, response format, temperature, top-p, top-k, and reasoning controls.
• Prompt and completion prices are both currently listed as $0 per million tokens.
• The model is categorized as a reasoning model aimed at coding and sustained agentic work.

These values can change because they describe a live preview, not a permanent product contract. The pricing tracker keeps a dated record rather than turning today’s zero into a promise about tomorrow.

One more confirmed detail matters for real work. OpenRouter’s Stealth page says prompts and completions are retained by the provider and are not used for training. “Not used for training” is narrower than “not retained.” Because the provider is anonymous, developers cannot complete the vendor review they might perform for a named API. Avoid secrets, credentials, unreleased source code, personal data, and client material unless your organization has independently accepted that risk.

What “stealth” means here

Stealth is a provider label, not a model architecture. It means the party operating the model has chosen to remain anonymous during this preview. It does not tell us the parameter count, training recipe, base model, lab location, safety process, or long-term business model. It also does not guarantee that an eventual reveal will map neatly to a currently rumored product.

Anonymous previews can be useful. A lab can collect real-world feedback without the brand expectations attached to a major launch. Developers can test a new capability before pricing and positioning are finalized. The cost is uncertainty: there is no named vendor roadmap, public system card, published service-level objective, or stable commercial contract to evaluate today.

That makes Ox Alpha more suitable for experiments than dependencies. If your application must work next month with a known support path, the anonymous preview creates a risk that raw capability cannot offset.

What the model appears designed to do

The million-token context window is the clearest practical differentiator. In principle, it lets a request carry more repository files, issue history, logs, specifications, images, or long documents at once than a smaller context window. More context is capacity, not judgment. A model can receive an entire repository and still focus on the wrong files, miss constraints, or produce an invalid patch. Developers should evaluate retrieval quality, instruction adherence, tool use, and verification behavior—not just how much material fits in the request.

Tool calling and reasoning controls make Ox Alpha compatible with agent frameworks that can inspect files, run commands, and iterate. Again, the harness matters. A raw single-attempt coding benchmark tests something different from a long-running coding agent that can read failures and repair its work. Our benchmark collection records the harness beside the score so those modes are not mixed into one ranking.

Multimodal input may be useful for reading screenshots, diagrams, and video frames alongside text. OpenRouter confirms the accepted modalities, but the anonymous provider has not published a modality-specific evaluation. Support for an image field is not evidence of accuracy on visual tasks. Test the exact screenshots or diagrams your workflow uses.

Who might be behind it?

The responsible answer is still: we do not know.

One community theory points to Z.ai or another model in the GLM family. Public investigators report tokenizer similarities, matching error strings, and similar outputs under controlled prompts. A detailed Reddit fingerprinting post lays out that argument. Those are black-box observations made by community members. They are interesting clues, but shared tokenization or infrastructure behavior does not prove ownership.

Another theory points to Google DeepMind, largely because social posts were interpreted as hints about a forthcoming Google model. The community discussion shows how that interpretation spread. No Google statement cited there names Ox Alpha, and timing plus suggestive language is weak attribution evidence.

Other guesses name different Chinese or US labs. None has produced an official statement tied to stealth/ox-alpha. The identity investigation tracks the theories, counterpoints, and the evidence that would be needed for confirmation.

What is not public

There is currently no provider-published model card, parameter count, architecture description, training-data summary, safety evaluation, benchmark suite, reliability commitment, permanent pricing, end-of-preview date, or broad commercial-use statement. There is also no independent leaderboard result that makes Ox Alpha directly comparable with every named model under the same harness.

That absence should shape the decision. It does not mean the model is bad; it means confidence should come from your own controlled evaluation rather than a reputation you cannot yet inspect.

Should developers try it?

Yes, if the test is low-risk and has a clear question. A useful evaluation might ask whether Ox Alpha can locate a bug across a large repository, follow an existing implementation plan, or interpret a UI screenshot while editing the related component. Define success first, use non-sensitive material, keep the same tool permissions for comparison models, and review the final diff and test output.

Do not treat “free” as “production-ready.” Availability, latency, identity, retention, and future cost remain separate decisions. Start with the OpenRouter setup guide, then compare your results with the public Ox Alpha benchmark evidence without assuming your harness will reproduce someone else’s score.

The simplest summary is this: Ox Alpha is a real, accessible, unusually large-context model with confirmed multimodal and tool-capable API metadata. Its creator and long-term product terms are not public, and its performance evidence is still early. That combination makes it worth measuring—and too uncertain to trust without measurement.

Last updated: 2026-08-23