Descriptions:
Sam Witteveen provides a hands-on look at Google’s newly released Gemini 3.1 Pro, the first model in the Gemini family to receive a point-release designation rather than simply a new preview date. The video arrives roughly 100 days after Gemini 3 Pro launched and examines what has changed — and why Google chose to call this a 3.1 rather than another incremental preview.
The benchmark numbers are striking: ARC-AGI jumps from 31% on Gemini 3 Pro to 77% on 3.1, and Humanities Last Exam shows a similarly large leap. Witteveen attributes much of this to what he describes as a ‘deep think mini’ capability baked into the new high-thinking mode, suggesting Google has transferred lessons from its Gemini Deep Think research models into the main Pro line. Notably, the model now supports three thinking levels — low, medium, and high — compared to just two on the previous version.
To demonstrate the difference in practice, Witteveen runs an International Math Olympiad problem through the model at high thinking, recording an 8-minute time-to-answer versus the 17-plus minutes previously required by the standalone Deep Think model — and getting the correct result. The same problem fails at low thinking, illustrating the tradeoff between speed and accuracy. Developers evaluating whether to upgrade from Gemini 3 Pro will find this a concrete, test-driven reference.
📺 Source: Sam Witteveen · Published February 19, 2026
🏷️ Format: Review
