Google's Gemini 3.5 Flash beats Pro on speed, but Pro holds reasoning edge

Google's latest model lineup introduces Gemini 3.5 Flash as a stable, production-ready model built for speed and agentic workflows, while Gemini 3.1 Pro Preview remains a preview model for the hardest reasoning tasks. Both models share a 1,048,576-token input window and 65,536-token output window, along with core features like function calling, code execution, structured outputs, thinking, and search grounding. However, Flash is explicitly designed for coding agent loops, tool-heavy functions, high-concurrency APIs, and batchable analysis—where latency directly impacts product quality. Google's launch benchmarks claim Flash outperforms Pro on Terminal-Bench 2.1, GDPval-AA, and MCP Atlas, though these are directional, not definitive for production.

For routing, the practical advice is clear: use Gemini 3.5 Flash as the first test for any new latency-sensitive workload, but do not delete Gemini 3.1 Pro Preview from reasoning-heavy or long-document routes until your own evals prove the switch is safe. Pro is a preview model with a separate custom-tools endpoint, making it suitable for cautious analysis, one-correct-answer tasks, and workloads where retries are costly. Both models lack support for audio/image generation, Live API, or Computer Use. Until production data confirms Flash covers all needs, route both models side by side—measuring wall-clock time, tool retries, router fallbacks, and manual review rates.