Gemini 3 Pro beats GPT‑5.2 on CAIS – 57.1 vision index

After last week’s GPT‑5.2 hype cycle, today’s third‑party evals paint a more nuanced frontier picture. On CAIS’s new Text Capabilities Index, Gemini 3 Pro edges out GPT‑5.2 with a 47.6 average vs 45.9, and a big lead in expert‑level reasoning (38.3 vs 29.9) plus Terminal‑Bench coding (53.4 vs 47.7). GPT‑5.2 still owns ARC‑AGI‑2 abstract reasoning at 43.3 vs Gemini’s 31.1, so the “hard math and puzzles → OpenAI” routing rule holds.

CAIS’s Vision Index tilts even harder toward Google: Gemini 3 Pro clocks a 57.1 average vs GPT‑5.2’s 52.6, with standout scores on embodied ERQA (70.2 vs 60.7) and MindCube spatial navigation (77.3 vs 61.7). OCR‑Arena tells a similar story for documents: GPT‑5.2 Medium lands #4 with Elo 1648 and 34.3s latency per page, behind Gemini 3 Preview, Gemini 2.5 Pro, and Opus 4.5 Medium.

WeirdML adds cost texture: GPT‑5.2‑xhigh nudges ahead in accuracy (~0.722 vs Gemini 3 Pro’s ~0.699) but burns around $2.05 per run against $0.526 for Gemini. Meanwhile, builders are openly dunking on SimpleBench after it ranks 5.2 below older models, treating it as noisy trivia rather than a routing oracle. Net: use GPT‑5.2 for peak abstract reasoning, but Gemini 3 Pro looks like the current default for multimodal coding, agents, and doc‑heavy workloads.