Ox Alpha is an anonymous large language model listed on OpenRouter as stealth/ox-alpha on August 20, 2026, offered free during an evaluation window. It advertises a 1,048,576-token context window and accepts text, image and video input. No lab has publicly claimed it, which is why it is called a stealth model — a common pattern where a vendor tests a pre-release model under a codename to gather real-world usage data before announcing it.

Is Ox Alpha really better than GPT-5.6 at coding?

Probably not, and the early claims did not hold up. First-wave independent tests reported 80% Pass@1 on the DeepSWE benchmark against 65% for Claude Fable 5 and 52% for GPT-5.6 Sol. By August 23, 2026 a fuller DeepSWE run by the same tester placed Ox Alpha roughly level with GPT-5.6 Sol mid rather than well ahead, and the LLM Coding Leaderboard scored it at #26 with 8.9 of 20 points on medium effort. No major public leaderboard has published a headline-confirming result.

Is it safe to use Ox Alpha on real code?

Treat it as unsafe for proprietary code. Nobody has confirmed who operates the model, which means nobody can tell you what happens to prompts you send it — retention period, training use, jurisdiction, or subprocessors. A free stealth endpoint is usually collecting data as the price of admission. Use it on public repositories, synthetic tasks or throwaway snippets only, and never route production traffic or customer data to an unattributed model.

Who is behind Ox Alpha?

Unknown as of August 25, 2026. Community analysis has focused on Chinese labs, partly because several have used anonymous OpenRouter listings as a launch pattern, but tokenizer and behavioural fingerprinting has not produced a confirmed attribution. Anyone stating the owner with certainty is speculating.

Quick Answer

What Is Ox Alpha? The Stealth Coding Model Explained

Published: August 25, 2026

The Short Answer

Ox Alpha is an anonymous model that appeared on OpenRouter as stealth/ox-alpha on August 20, 2026, free to use, with a 1,048,576-token context window and multimodal input covering text, images and video. It generated a week of coding-benchmark headlines and an industry guessing game about which lab built it.

As of August 25, 2026: nobody has claimed it, and the benchmark claims that made it famous have already softened.

What Is a Stealth Model?

Stealth listings are a normal pre-release tactic. A lab pushes an unannounced checkpoint to a public router under a codename, prices it at zero, and watches how it performs on real developer traffic before committing to a launch. The lab gets free evaluation at scale plus a viral moment if the model is good. Recent frontier releases from several vendors were tested this way first.

The trade the user makes is implicit: you get a frontier-class model for free, and the operator gets your prompts.

The Benchmark Story — and the Correction

This is the part worth reading carefully, because it is a clean case study in how AI benchmark claims propagate.

Stage	Claim	Date
First independent tests	80% Pass@1 on DeepSWE, vs 65% Claude Fable 5 and 52% GPT-5.6 Sol	~Aug 21, 2026
Media pickup	”Beats GPT-5.6 and Claude Fable 5 at coding”	Aug 21–22, 2026
Fuller re-run by same tester	Roughly level with GPT-5.6 Sol mid	Aug 23, 2026
LLM Coding Leaderboard (OpenCode)	#26, 8.9 / 20 points, medium effort	Aug 23, 2026
Major public leaderboards	No confirming result published	As of Aug 25, 2026

The first number came from a partial benchmark subset. The corrected number came from the complete set. The headlines were written from the first number and largely not updated — which is the ordinary lifecycle of a benchmark claim in 2026.

A single partial-subset score is not evidence. It is a hypothesis that a complete run either confirms or, as here, deflates.

What Ox Alpha Actually Offers

Stripped of the hype, the concrete specifications are still notable:

1,048,576-token context — genuine frontier-tier context length, matching Claude Opus 5 and DeepSeek V4
Multimodal input — text, images and video
Free during the evaluation window — reported as roughly a week
Competent, not leading, at coding — mid-pack on a full evaluation

A free model with a 1M context that codes about as well as a mid-effort frontier configuration is genuinely useful for the right task. It is not a frontier-beating breakthrough.

The Real Story Is Data, Not Benchmarks

The question developers should be asking is not “how good is it” but “where do my prompts go?”

With a named provider you get a data processing agreement, a published retention window, a stated position on training use, a jurisdiction, and a company that can be sued. With stealth/ox-alpha you get none of these, because there is no identified counterparty. OpenRouter routes the request; it does not underwrite the operator’s data practices.

Practical rules if you want to try it:

Public code only. Open-source repos, sample projects, synthetic tasks.
No secrets, ever. No keys, no customer records, no internal architecture.
No production traffic. A model that can vanish in a week is not a dependency.
Assume prompts are retained and used for training. That is the most probable price of a free frontier endpoint.

Should You Care?

If you are benchmarking models for a build decision: wait for attribution. An unclaimed model cannot be procured, cannot be contracted, and cannot be relied upon. If it is a real pre-release, it will get a name and a price sheet within weeks, and that is the version worth evaluating.

If you are curious: use it on a public repository for an afternoon and form your own view. Your own task-level evaluation is worth more than the entire benchmark discourse around it — a point that applies well beyond this model.