What is Ox Alpha? Everything we know about the stealth model

Last updated: August 25, 2026. Ox Alpha is still in its stealth preview and the story is moving. I'll update this article as new evidence lands and again the day the identity is confirmed.

On August 20, 2026 a new model appeared on OpenRouter under the ID stealth/ox-alpha: a million-token context window, image and video input, a $0 price tag and no lab name attached. Five days later it has swallowed billions of tokens from coding agents and earned a "very impressive" from Stripe CEO Patrick Collison. It has also kicked off the best guessing game the AI community has had all year. Nobody has claimed it. The evidence, though, points somewhere specific.

I've spent the past few days reading every fingerprint analysis, benchmark run, terms page and Reddit thread I could find on this thing. Here's what's verified and what's still speculation, plus where I stand.

Ox Alpha specs and pricing

Everything below comes from the official OpenRouter listing, which is the only first-party surface this model has:

• Model ID: stealth/ox-alpha, served by a single provider labeled "Stealth"
• Context window: 1,048,576 tokens, with a 131,072-token max output
• Input: text, images and video; output is text only
• Pricing: $0 per million tokens, in and out, for the duration of the preview
• Features: tool calling and JSON output via response_format; a reasoning_effort parameter (low, high, max) that defaults to max

The listing describes a reasoning model aimed at coding and sustained agentic work, and it's blunt about the mystery: the model "is developed and operated by a third-party provider who has chosen to remain anonymous during this preview". OpenRouter routes the traffic and states plainly that it isn't the developer.

The same model also runs on OpenCode's Zen and Go routes, where the team announced a free week starting August 20 and claimed capacity for 100 trillion tokens per day. That free window closes around August 27, so if you want to test it at zero cost, this is the week.

Who is behind Ox Alpha?

Short answer: nobody has confirmed anything, and as of August 25 no lab has stepped forward. Longer answer: the technical evidence converged fast, and it didn't converge on Google.

The GLM-5.3 fingerprint evidence

Within 48 hours of launch, independent testers published separate fingerprints that all point at Zhipu (Z.ai), the Chinese lab behind the GLM family.

The tokenizer test is the strongest one. Probe prompts mixing languages, emoji, code and SQL produce token counts that match GLM-5.3 on every single probe, 11 out of 11 in the set collected by DigitalApplied's stealth model census, while no other lab's model matched more than four. A tokenizer is about as close to a fingerprint as this field gets. Two unrelated model families don't match on every probe by accident.

Then there's the serving stack. A leaked stack trace from a malformed request reportedly contained a Java class path matching Zhipu's documented cloud API internals, and the error codes returned on bad requests follow the GLM API's dialect. A separate analysis found that Ox Alpha consumes video tokens the same way GLM-5V-Turbo does, down to frame sampling and resolution scaling. Prediction markets have priced all this in: Manifold bettors currently put Zhipu at roughly 90%.

None of it is confirmation. A fingerprint proves shared tooling and infrastructure, and the 1M context doesn't match anything Zhipu currently sells, which suggests a variant or an unreleased successor rather than a straight rebrand. Still, if I had to bet, my money is on GLM.

The Gemini and Microsoft MAI theories

Early chatter guessed Google, mostly on timing and feel (Gemini rumors were in the air the same week). The fingerprint work argues against it: Ox Alpha reportedly failed a vision test Gemini models pass, and none of the serving signals look like Google infrastructure. TechCrunch's report also mentions the theory that this is an unreleased Microsoft MAI model, a claim that arrived without technical evidence and has stayed that way. Both remain possible in the way most unfalsified things are possible.

Ox Alpha benchmark results: The 80% claim and the real score

Here's where the hype ran ahead of the data, and where a lot of the day-one coverage is now simply wrong.

The number that went viral was an 80% Pass@1 on DeepSWE, from a 10-task sample someone ran on launch day. Ten tasks. A coin-flip-sized sample, repeated across half the internet as "mystery model beats every frontier lab".

The full run tells a calmer story. StealthModelWatch's complete 113-task DeepSWE run scored 58.4% Pass@1 (66 of 113 solved, 95% CI of 49.2% to 67.1%), which the tester places roughly level with Claude Opus 4.8. Their report is direct about the earlier figure: "the rumored ~80% pass rate is completely incorrect." Two details from the same data are worth keeping. Around 80% of tasks passed at least 90% of their tests, so the binary score hides a lot of near-misses. And 11 of the 47 failures were formatting errors, responses that came back without tool calls, rather than reasoning failures.

My read: the model is frontier-class at coding without being the leaderboard anomaly the early posts described. And 58.4% from an anonymous free preview is the impressive number. It didn't need inflating.

Ox Alpha speed and latency

The consistent complaint is speed. Public telemetry has hovered in the mid-20s tokens per second at the median, with a few seconds before the first token and P95 tails that stretch much longer at peak hours. That full benchmark run took over 20 hours for 113 tasks. Reasoning runs before every answer and defaults to max effort, so anything interactive feels sluggish. For agents grinding overnight, nobody cares. For a chat loop, you will.

Ox Alpha data retention and the stealth model terms

This is the section most explainers skip and the one I'd read first.

OpenRouter's stealth program has a default deal: the model is free and the anonymous provider can collect your prompts and train on them. That's the exchange, spelled out in the Stealth Model Terms. Ox Alpha carries an exception, the only one among all 14 stealth models to date: prompts and completions "are retained by the provider and are not used for training".

Read that twice. No training is a real concession. Retention still means every prompt you send sits in the logs of a company that won't tell you its name, under a policy you can't read. OpenCode advertises zero retention on its own routes, which helps on their side of the wire and changes nothing upstream. Working at a hosting provider means I read data-handling clauses for a living, and my rule here is the one I'd apply to any anonymous endpoint: treat every prompt as public. Throwaway branches and open-source code are fine. Proprietary code and customer data are not.

Is oxalpha.com the official Ox Alpha website?

No. Neither oxalpha.com nor oxalpha.io belongs to the lab behind the model. Both are third-party front-ends that appeared within days of the launch to catch the search traffic, and to their credit both say so. Oxalpha.com's own about page states it's "not affiliated with, endorsed by, or operated by whoever built the model". The real provider has no website and no social presence anywhere. The OpenRouter listing is the only official surface, so treat any oxalpha-something domain as an unofficial wrapper that adds its own terms and its own logging between you and the model.

How to access Ox Alpha

Two legitimate routes. On OpenRouter, pick stealth/ox-alpha in the chat playground or call the OpenAI-compatible API with your key; it speaks the standard chat completions format, so anything that talks to OpenRouter can use it, and agent harnesses from Claude Code to Hermes Agent already push billions of tokens through it daily. On OpenCode, it shows up as "Ox Alpha Free" on the Zen and Go routes for the length of the preview.

Set expectations before wiring it into anything real. Rate limits are shared and undocumented, and the endpoint can vanish the day the preview ends. The speed is what it is. If you're collecting free options for agent setups more broadly, our guide to free AI models covers ones that stick around after the hype week.

When will Ox Alpha's identity be revealed?

Stealth models get revealed; that's the pattern. Since April 2025, OpenRouter has listed 14 of them and half were later claimed officially by their labs. Quasar Alpha and Optimus Alpha ran under two weeks before OpenAI confirmed both as GPT-4.1 test builds. Horizon Alpha lasted about a week and turned out to be an early checkpoint of the GPT-5 family, the same family we covered in our GPT-5 guide last summer. Others were never claimed and quietly went offline when their test window closed.

So the realistic window for a reveal, if one comes, is days to a few weeks, and the free period ending around August 27 is a natural moment for it. Names move fast in this space anyway. We traced OpenClaw's two rebrands in a single month once, and that story still gets referenced today for the same reason this one will: somebody has to write the timeline down while the links are live.

When the confirmation lands, I'll update this article the same day. If the fingerprints are right, a Chinese lab just put a frontier coding model in front of a few hundred thousand Western developers for a free week, without saying a single word about itself. That silence has outperformed most launch posts I've read this year.