ox-alpha

Revealed as an early version of GLM-5.3-Flash. The anonymous “stealth” model that ran free on OpenRouter (stealth/ox-alpha) from Aug 20, 2026 was confirmed by Z.ai as an early build of GLM-5.3-Flash, which the lab shipped officially on Aug 26 - and it was served entirely on Chinese AI chips. Z.ai’s Zixuan Li confirmed ox-alpha was an early version, with the official release delivering stronger performance and significantly better stability. See the GLM-5.3-Flash card for the real model; this row is superseded by it.

Why it mattered: Z.ai tested GLM-5.3-Flash anonymously to gather real user feedback before naming it. It quickly became the most popular model of the week - the biggest OpenRouter/OpenCode launch to date. OpenCode reported 42 trillion tokens served in just 6 days, making it the most-used model after DeepSeek Flash’s 56-day run - all of that traffic served on Chinese-made AI accelerators.

Agents on Rails benchmark (Aug 2026, Le Mans round). 82.5% accuracy on 63 runs - tied with Grok 4.6 and behind only the Opus 5 / Kimi K3 / Fable 5 cluster. API recall 28.6%. Slow (19m 26s median) but strong. Superseded by the stronger, more stable official GLM-5.3-Flash release.