Google has released Gemini 3.1 Flash-Lite, a superfast new model that excels at visual tasks. Robin Sloan, a developer who weaves these models programmatically into other systems, calls Gemini his favorite for speed, price, and versatility. While competitors like Anthropic and OpenAI focus heavily on coding agents, Google continues expanding what Sloan describes as a flexible, dare I say general, intelligence.

This latest release comes with a caveat. Google recently deprecated Gemini 3 Pro after a notably brief existence, joining other companies in the troubling practice of yanking away models developers depend on. Even when replacements benchmark better, they can contain unpredictable regressions along subtle dimensions. Sloan notes this issue is completely resolved with self-hosted models, though none currently match Gemini's visual acuity.

The visual capabilities remain Gemini's standout feature. For developers building systems that process images or video, the combination of speed and accuracy makes Gemini particularly compelling. While the model ecosystem continues evolving rapidly, Google's approach of maintaining broad capability rather than specializing in specific tasks like coding represents a distinct strategic bet.