Early tests of mystery model Ox Alpha can't agree if it's frontier
The mystery model that appeared on OpenRouter this week with no name attached is still free, still anonymous, and now the center of a very public argument about whether it's any good. Early benchmark results are pointing in opposite directions — and the fight over Ox Alpha's report card is becoming the story.
Early tests of mystery model Ox Alpha can't agree if it's frontier. The anonymous model surfaced on OpenRouter earlier this week offering a 1M-token multimodal context window, zero data retention, and what its distributor calls capacity for 100 trillion tokens a day — free for roughly a week, with integrations into coding tools like Codex and OpenCode following within days. The first scores were spectacular: developers reported 80%+ on a DeepSWE subset against GPT-5.6 Sol at 52% and Claude Fable at 65%. Since then the verdicts have fractured. A cybersecurity benchmark team scored it above one frontier model but below most of the rest of the frontier; an independent evaluator ranked it alongside models two generations old; a DeepMind researcher called it average at coding; and others found it mediocre at aerial-image analysis and weaker than Opus 5 on a build-off. Nobody is accusing anyone of bad numbers — the likelier explanation is that a long-context agent model grades differently depending on which slice you test.
Identity sleuthing has moved past the GLM consensus. Yesterday's leading theory — tokenizer fingerprints matching Zhipu's GLM family — now has company. One developer measured its serving infrastructure as US-hosted, which sits awkwardly with the Chinese-lab theories, and Google employees' very public enthusiasm for the model keeps feeding a "secret Gemini" rumor the evidence doesn't clearly support. The honest answer remains: nobody has claimed it, and every fingerprint so far is circumstantial.
The distribution play may be the real headline. In under 48 hours the same anonymous endpoint went from a single stealth launch to free availability across OpenRouter, Codex, OpenCode, and — per one observer's tally — a portal advertising a quadrillion tokens a day. That is tens of millions of dollars in inference spend being given away by a lab nobody can name, which either signals a capability stress-test ahead of a paid launch or the most expensive marketing stunt the model economy has produced. Either way, the free window is closing: usage stops counting in about six days, and when the meter starts running, we'll finally learn who built this thing and what it costs.
What to watch: whether any lab claims Ox Alpha before the free week ends — and whether the first paid API price reveals which benchmarks were measuring it fairly.
If you've run Ox Alpha yourself, which camp are you in — frontier-grade or overhyped? Tell us in the comments.
