Ox Alpha's viral 80% coding score used just 8 tasks

A stealth model called Ox Alpha appeared free on OpenRouter on August 20. It carries 1,048,576 tokens of context, a 131,072-token max output, and takes text, image and video as input.

The number that made it go viral doesn't hold up. The 80% coding score covered only 8 hand-picked tasks on the DeepSWE benchmark. Once the full 113-task run finished, the model resolved 58.4%, a 22-point drop from what got shared around.

Stripe's Patrick Collison called it "very impressive." No lab has stepped forward to claim it, and the guessing splits three ways: Z.ai's GLM family, Microsoft's MAI, or Xiaomi's MiMo, per TechCrunch's report.

TechCrunch's report doesn't say when the 8-task number was shared relative to the full run finishing, so there's no way to tell if it was an early preview or a number picked to look better than the real result. Builders testing free frontier-scale models on OpenRouter should expect the same pattern: a headline score arrives first, and the full evaluation number arrives later, with less attention attached to it.