Mysterious AI Model Ox Alpha Beats Claude Fable and GPT-5.6 on Coding Benchmarks

A mysterious AI model called Ox Alpha has appeared on OpenRouter with no company name attached, and it is already topping major coding benchmarks.

What We Know

  • Released August 20 on OpenRouter, OpenCode, Cline, and Nous Research
  • Free to use during preview period
  • 1 million token context window
  • Accepts text, image, and video input
  • Supports tool and function calling

Benchmark Performance

On DeepSWE, a benchmark measuring how often a coding agent resolves real GitHub issues correctly:

  • Ox Alpha: 80 percent on initial 10-task sample
  • Claude Fable 5: 65 percent
  • GPT-5.6 Sol: 52 percent

However, full 113-task runs showed Ox Alpha at roughly 63 percent, level with GPT-5.6 Sol.

Who Built It?

The most popular theory points to Zhipu AI model family. Independent fingerprinting matched Ox Alpha tokenizer to GLM-5.3 on every test, and its video-token spend matches GLM-5V-Turbo exactly.

Stripe CEO Patrick Collison called it very impressive, helping push the model into wider conversation.

Source: Decrypt, August 2026