February 19, 2026
Version bump, drama jump
Gemini 3.1 Pro Preview
Google sneaks out Gemini 3.1 Pro—bigger brain, messy name vibes
TLDR: Google’s Gemini 3.1 Pro Preview quietly dropped with a giant memory and a new Medium “thinking” level. Commenters split: some roll eyes at the minor version bump and messy naming, while others tout huge benchmark gains; many say they’ll stick with cheaper, faster Flash—at least for now.
Google quietly slid the Gemini 3.1 Pro Preview into the world overnight, a “thinking” AI with a massive memory (it can digest huge documents) and new thinking levels—Low, Medium, High—plus agent-y tricks like search, code execution, and structured outputs. The official blog post exists, but commenters swear it appeared in Vertex AI at 2am like a surprise mixtape.
Cue the drama: Topfi threw shade at the version bump, grumbling that the only real change is “Medium reasoning,” and that Google’s model names “have long gone from even trying to make sense.” Meanwhile, boosters fired back with numbers: sigmar flexed huge benchmark jumps, claiming arc-agi-2 leaped from 31.1% to 77.1% and agent tests doubled. WarmWash added to the spectacle, calling the rollout disjointed and predicting a proper announcement later.
Pragmatists shrugged. mark_l_watson said he’ll keep using Gemini 3 Flash because it’s cheap and fast, while DIYers cheered small models on home rigs. The memes wrote themselves: “AI version bingo,” “Medium thinking is the new latte size,” and “woke up to a 2am brain upgrade.” Despite stricter rules (don’t mix old and new settings) and no tuning yet, early tests look promising—if you can survive the naming chaos.
Key Points
- •Gemini 3.1 Pro Preview is a multimodal agentic and coding model with improved complex instruction following and output efficiency over Gemini 2.5 Pro.
- •It supports audio, image, video, text, and PDF inputs, text outputs, a 1M-token input limit, and a 64k-token output limit.
- •Thinking mode now includes low, medium, and high levels; medium reasoning is newly added compared to Gemini 3 Pro Preview.
- •Features include Google Search grounding, Vertex AI RAG Engine, URL Context, Code Execution, Structured Output, caching, Batch Prediction, and Provisioned Throughput; tuning is not available.
- •Stricter validation and API changes include replacing thinking_budget with thinking_level, disallowing both together (400 error), and updating usage metadata and media resolution defaults.