The stealth model "Ox Alpha" is unmasked as GLM-5.3-Flash

What's this about?

The anonymous model that appeared on OpenRouter as "ox-alpha" — which this digest covered as a mystery on 21 August, and which Ethan Mollick was unimpressed by — is confirmed as Z.ai's GLM-5.3-Flash. The weights are published. Per oran_ge's breakdown: 320B parameters with 18B active, priced at roughly a fortieth of Opus 4.8, scoring 57 on the Artificial Analysis Intelligence Index (level with Opus 4.8), and — the detail that got the most attention — all of its traffic, reportedly over 100 trillion tokens a day, is served on Chinese-made chips.

What exactly is a "stealth model"?

Labs sometimes release a model anonymously on a routing service first, under a codename, so people evaluate it without knowing whose it is or what it cost. It produces unbiased first impressions. The tactic worked here in an unusual way: sceptical early reviews came from people who assumed a small experiment, and the reveal reframes those same results as coming from a model priced at a fortieth of the frontier.

What's the community saying?

1. The most-replied comment is a warning to read Z.ai's terms of service — a broad perpetual licence over your inputs, outputs, name and profile picture, plus vague prohibitions on content harming Z.ai's "interests" or any country's "national interests".
2. "RIP Nvidia shareholders" — the Chinese-chip detail is read by many as the genuinely significant part, above any benchmark.
3. Contradictory measurements are openly flagged: it ranks below GPT-5.4 Nano on LiveBench while beating Fable substantially on the vendor's own site.
4. A detailed hands-on report: used via OpenRouter for days of coding, placed between Sonnet and Opus, few mistakes but "not that smart", and it degraded into doom loops several times.
5. A recurring meta-point that Chinese labs are so used to inflating benchmarks that this announcement undersells a model that is actually good.
6. Scepticism about the ecosystem rather than the model — Z Code looks like a Codex fork, as does Kimi's, and the harnesses feel like an afterthought.