leejk/ jk lee
Models: lining up the frontier base models

Landscape · 2026.06

Google — Gemini Models

Google's base models, gathered in one place

·
#models#google#gemini#landscape

A vendor page collecting Google's base models (the Gemini line). Per-model strengths, weaknesses, and operational constraints are sourced in each standalone piece; this page only sorts which model takes which job.

TL;DR. A vendor page gathering Google's base models. Per-model strengths, benchmarks, and operational constraints are backed by primary sources in each standalone piece below; here we only classify which model takes which job.

What this slot is

The set of base models Google ships for general use. It is one vendor column in the models slot that coding agents, frameworks, and skills ride on — and even within one vendor the job splits per model: reasoning/science (the 3 Pro line), agentic/coding (3.5), generation (Omni).

Lineup — a generation just turned over

Google's base models span two generations. Gemini 3.1 Pro (2026-02-19) is the reasoning/science core that holds this blog's models table science-reasoning #1, and Gemini 3.5 (2026-05-19, I/O) turned the generation over with "frontier intelligence with action," taking agentic/coding (3.5 Flash shipped first — Terminal-Bench 2.1 76.2%, GDPval-AA 1656, MCP Atlas 83.6%). Separately, Gemini Omni takes omnimodal generation (video and more).

But the turnover has stalled halfway (as of 2026-08). Google previewed 3.5 Pro at I/O as shipping "next month" and it still has not landed — on 2026-07-21 Google released three more Gemini models and 3.5 Pro was not among them. Reporting puts it months behind while coding, math, and hallucination reliability get fixed. So the current Pro-tier model is still 3.1 Pro — Flash took agentic and coding while the previous generation holds the Pro slot, leaving the lineup out of joint.

Models

Each model's strengths and weaknesses are sourced in its standalone piece.

  • Gemini 3.1 Pro — the reasoning/science core (ARC-AGI-2 77.1% verified, GPQA science-reasoning #1). But it trails GPT-5.5 on agentic/coding, and its own successor 3.5 Flash is now overtaking it there too.

Sources

Sub-documents