March 3, 2026
New Gemini model out today: 3.1 Flash-Lite, superfast and very capable.
The Gemini models remain my favorites: for their speed, price, and versatility, especially in visual tasks. Keep in mind that I am weaving these models programmatically into other systems, rather than yapping with them — although I think Gemini is a fine yapper, too.
As Anthropic and OpenAI lean HARD into the coding agent thing — tuning their models heavily for the task — Google continues to expand and refine a flexible … dare I say general … intelligence.
P.S. Although, let me add, the quick deprecation of Gemini 3 Pro is annoying and irresponsible. This isn’t just a Google thing — although this model had a notably mayfly existence — and there is really no telling if or when any model you depend on will be yanked away. Even if the replacement is “better” according to many benchmarks, it might also contain unpredictable regressions along subtle dimensions. (You can browse the responses to Google’s deprecation announcement for examples.)
This issue is totally resolved with a self-hosted model, of course, but there aren’t yet any self-hosted models with Gemini’s visual acuity. I do expect that to change before too long!