On 20 August 2026, an anonymous model named Ox Alpha appeared on the OpenRouter and OpenCode platforms, with the provider simply listed as stealth. Its specs: 1,048,576 tokens of context, output capped at around 131,000 tokens, text, image and video input, function calling, and a price of zero for about a week. No lab has claimed it. Its system prompt reportedly tells it to present itself as developed by an undisclosed organisation.
This is the strangest story of the week, and probably the most interesting to pick apart, because it involves two things we rarely talk about: how these models are actually tested, and what you give away when you don't pay.
What a stealth release is
Let's start with the practice, because it explains everything else. A lab preparing a model needs to expose it to real-world usage before naming it. Internal tests, even massive ones, don't reproduce what thousands of developers do with a tool when they're trying to solve their own problems.
The solution is to publish the model under a codename, without claiming authorship, and make it free for a short period. Users flock in, push considerable volumes of tokens through it, and the lab observes. This would be the fifth release of this kind in about six months.
The anonymity has genuine methodological value: nobody judges the model through the reputation of its brand. It's a form of blind testing, comparable to what we noted in the video generation rankings, where votes are cast without knowing which model produced what. Incidentally, if the model disappoints, nobody has anything to lose.
What is established
Let's carefully separate verifiable facts from hypotheses, because this case is saturated with speculation.
The technical specs appear on the platform's official listing, which describes a reasoning model designed for code, long-running agentic work and production workloads.
The free access and its end date. Open access on OpenRouter ran until 24 August, and a parallel offering on a command-line agent was announced for about a week from 20 August, i.e. until the 27th.
The anonymity. The platform explicitly states that the model is developed and operated by a third-party provider that has chosen to remain anonymous during the preview. It routes traffic; it is not the designer.
Who is behind it?
The community has engaged in a fairly remarkable piece of detective work: analysing the tokenizer, error messages, and the language of the server's technical logs.
The dominant hypothesis points to the GLM family from Chinese lab Z.ai. It's consistent with the timeline, as that lab announced GLM-5.3 a few days earlier. There's also precedent: a stealth model that appeared in February was later attributed to that same family.
But let's be clear: nothing is confirmed. Previous stealth releases have turned out to belong to labs other than those the community suspected. Any claim about Ox Alpha's identity is, at this stage, speculation, however elaborate.
The performance claims
One claim has circulated widely: a score of 80% on a coding test, versus 65% for Claude and 52% for GPT-5.6. It comes from a user trial covering around ten tasks, not an audited benchmark.
Ten tasks allow no conclusions. That's the very principle we set out in our article on benchmarks: a tiny sample produces gaps that are largely down to chance. The model impresses experienced users, which is a real signal. It is not a measurement.
Serving a model of this size costs money. We explained in our article on the two costs of AI that inference is a cost incurred on every request. An actor offering near-unlimited access for a week is taking on a considerable bill. So: what do they get in return? Visibility, which is legitimate. Usage feedback on real tasks, which is valuable. And potentially data. Thousands of developers have pushed entire repositories through this model in a few days. A repository contains credentials, access keys, customer data, licensed code. Nobody knows who receives all of that.
What free access does not transfer
There's a useful formula, taken from a guide aimed at technical teams: a bill of zero euros does not transfer the authorisation to disclose assets.
The fact that a service is free does not absolve you of your obligations to your clients, employers or partners. If your repository contains data covered by a confidentiality agreement, sending it to an anonymous provider is a contractual problem, regardless of the price.
That's exactly the reasoning in our selection guide: the first filter is never performance, it's whether your data can leave and under what right.
What to do, concretely
Use it as a disposable evaluation path, on sanitised tasks: sample code, a public project, a problem rebuilt without real data. It's a rare chance to test cutting-edge capability without a budget.
Build no production dependency on it. Anonymous, free, time-limited access is not a service: no availability commitment, no point of contact, no recourse. We've seen what a model's disappearance costs even with an identified, contractually committed provider.
Treat everything you send as public. That's the only prudent assumption in the absence of verifiable information about retention.
What this practice reveals
Three lessons go beyond the curiosity of the moment.
Real-world testing has become the scarce resource. Labs have compute and data, but not authentic usage at scale before launch. Offering a model free for a week amounts to buying millions of hours of real usage, paid in visibility rather than money.
Brand matters less than you'd think. That developers push billions of tokens through a model whose origin they don't know shows that the selection criterion has become observed performance on their own tasks, not the provider's reputation.
Cutting-edge capabilities are becoming commoditised. A model of this level can now appear without a name, be adopted en masse within forty-eight hours, and get coverage in the international financial press without anyone knowing who built it.
What to remember
Ox Alpha is an interesting opportunity and a potential trap, and both are true at the same time. Interesting, because testing a possibly leading model for free, on sanitised tasks, provides information few people can otherwise afford. A trap, if you make it anything other than a disposable test bench.
Finally, there's something quite revealing about the fact that part of the industry now tests its products by making them unrecognisable. That suggests those who design them know perfectly well that the name on the label influences judgement, including that of professionals who believe they're immune to the effect.
And it recalls an old rule, which the novelty of the object sometimes makes us forget. When an expensive service is offered free and without an identity, the question is not whether there's a catch. It's what is being exchanged, and whether you're in a position to exchange it.