#ScreenSpotPro
📊 Pro results: 93.3% #ScreenSpotPro, 91.0% OSWorld-G, 88.6% #OSWorld-Verified, 73.9% #WebArena-Verified. That beats trillion-parameter Qwen3.8-Max and Kimi-K3 on all four. Mini hits 85.3% on OSWorld-Verified with 3B active parameters
November 5, 2026 at 9:22 PM
The screenspotpro benchmark suggests computer use soon 👀
November 19, 2025 at 3:53 AM
Yeah, like 4,9% to 31,1% on ARC-AGI-2, or 0,5% to 23,4% on MathArena Apex (where neither Gemini 2,5, Claude 4,5 sonnet, nor GPT 5,1 got more than 1,6%). Or ScreenSpotPro, going from 11,4% (best previous model: Claude at 32,6%) to 72,7%. Love the MRCR score especially. Really impressive stuff.
November 18, 2025 at 3:03 PM