Stealth Model 'Ox Alpha' Matches GPT-5.6 in Context Arena Tests

scaling01 · x · 2026-08-24

A new stealth model named "Ox Alpha" has been added to the Context Arena, running via OpenRouter. It achieves a 75.3% AUC@128k score on high thinking settings, matching Gemini 3.1 Pro and GPT-5.6 Luna while beating Claude Sonnet 5. Interestingly, enabling max thinking reduces accuracy beyond 32k context length despite using 34% more tokens. The model maintains a 39.0% score at 512k context length.

Original post →

More from Models

Models channel →