FULL STORY

Ox Alpha: Anonymous Mystery Model Sparks Speculation

An anonymous model dubbed Ox Alpha quietly appeared on OpenRouter with no vendor attribution, fueling speculation it was a new Zhipu GLM release. Cola later announced its Ox model claiming Terra-level capability, advancing the mystery.

2026-08-21 ~ 2026-08-22 · 2 episodes · 39 posts

Episode 1 · Mystery Model Ox Alpha Appears, Suspected Zhipu GLM Release (2026-08-21, 37 posts)

Around August 21, an anonymous model named Ox Alpha quietly appeared on OpenRouter and OpenCode—no vendor attribution, no paper, no announcement. It offers a 1 million token context window, multimodal input across text/image/video, nearly a week of free access, and claims zero data retention. Multiple testers have reported stunning performance, and the community widely suspects, based on black-box fingerprinting tests, that it's built on the Zhipu GLM-5.3 architecture, though as of this writing no official claim has been made and its identity remains a mystery.

Confirmed

  • Ox Alpha is available on OpenRouter and OpenCode, with a 1M context window and multimodal input; currently free with nearly unlimited quota, and OpenCode claims compute capacity of 100 trillion tokens processed per day
  • Scored roughly 80% on the DeepSWE 10-task benchmark, beating GPT-5.6-sol (52%) and Fable (65%)
  • @Teknium tested OX Alpha Max in Hermes Bot and called its frontend quality beyond expectations—abnormally intelligent
  • @FlunkyGraphics ran three black-box fingerprint tests (tokenizer consistency, etc.) and confirmed Ox Alpha is based on the GLM-5.3 architecture
  • @omarsar0 (Elvis) said an informed source confirmed the model's true identity to him (not yet public), describing it as spectacular and saying it will change the expected timeline of the multimodal agent race

Unconfirmed

  • The model's owner: the prevailing guess is Zhipu—possibly a GLM-5.3 vision variant, GLM 5.3 Flash, a faster small variant, or even something to be named GLM-6 (a leak relayed by @hunarbatra); @teortaxesTex inferred from speed, TTFT, cache hit rate, and other measurements that it belongs to the GLM family rather than DeepSeek, and said that if it turns out to be a non-Chinese model (e.g., Grok), a reassessment would be needed
  • @rickasaurus relayed that the model claims to be Mistral's Le Chaton Fat, which contradicts other evidence and is of questionable credibility
  • @trollkhan noted that the claim of SWE scores surpassing Fable matches m9's data, but these are all community tests rather than official benchmarks

Why it matters

If the fingerprint tests hold up, this would be yet another anonymous surprise release from Zhipu's GLM family showcasing frontier coding and agent capabilities. According to @hunarbatra's relayed information, its SWE/Cyber benchmark performance has reached mythical levels—outperforming existing frontier models—and could reshape the multimodal agent landscape; the free, unlimited access also lets developers test and verify it first-hand.

17 more related posts →

Episode 2 · Cola Launches Ox Model Claiming Terra-Level Capability at Flash Speed (2026-08-21, 2 posts)

Cola released its Ox model, claiming multimodal, small-parameter, strongly post-trained performance at Terra level with Flash-level speed, offered free for a limited time; developer tests report solid technical strength.