Ox Alpha Is Free, Frontier-Class, and Unclaimed

On August 20 a model called Ox Alpha appeared on OpenRouter and in the OpenCode client, listed under the provider name "stealth". Its card describes a reasoning model built for coding, sustained agentic work and production workloads. It carries a 1,048,576-token context window, accepts text, images and video, and will return up to roughly 131,000 tokens in a single completion. It costs nothing to use. Nobody has said who made it.

That combination is why the listing spread as fast as it did. A million-token context and video input put Ox Alpha in the same bracket as the frontier systems that cost real money per million tokens, and developers who ran it against their own agent harnesses came away describing it as genuinely competitive rather than a curiosity. Stripe chief executive Patrick Collison called the model "very impressive" — a notable endorsement, though one worth reading alongside the fact that Stripe acquired OpenRouter for $7 billion earlier this month.

OpenRouter has been precise about what it is and is not. The company says it routes requests to Ox Alpha but is not the model's developer, owner or provider, and that the third party behind it has chosen to remain anonymous for the duration of the preview. The data terms differ by route: OpenCode says its paths retain nothing and train on nothing, while OpenRouter says the unnamed provider does retain request data but does not train on it. Access on OpenRouter requires an account and an API key; OpenCode serves it through its free Zen gateway tier and its paid Go plan. OpenCode has advertised aggregate capacity of 100 trillion tokens a day, a service-wide figure rather than a per-user allowance.

The guessing has mostly pointed at China. Tokenizer probing and behavioural fingerprinting led several developers toward Z.ai's GLM family, which shipped GLM-5.3 earlier in August and has the coding pedigree to match. The theory is not airtight: Ox Alpha takes image and video input, and GLM-5.3 is text-only. Competing threads have floated an unreleased Microsoft MAI checkpoint, a Phi derivative, and an IBM Granite lineage, with roughly the confidence you would expect from evidence assembled out of sampling quirks and refusal styles. No lab has stepped forward.

Anonymous previews have become a recognisable pre-launch tactic — a way to collect unbiased evaluations and real agentic traffic before a name and a price tag anchor expectations. What makes this one unusual is the scale of the giveaway. Serving a million-token context free to an open gateway is expensive, and whoever is paying that bill is buying something specific: comparison data gathered before anyone knew whose model they were grading.

The preview is reportedly a roughly one-week window, which would put the cutoff around August 27. Until it closes, the most interesting artefact here is not the benchmark score but the experiment itself — a frontier-tier model evaluated entirely on its output, with the brand stripped off.