Gemini CLI: Models & Context Windows

Gemini CLI runs in auto mode by default: it classifies each prompt as simple or complex and routes it to a Flash or Pro model on its own. This guide covers the model lineup you can reach, how to pin or switch models, how auto-routing decides, and how your auth method sets your daily quota. After it you can select a specific model, read which one actually ran, and budget your requests.

All current Gemini coding models share a ~1M-token input window, so model choice is about speed, cost, and reasoning depth, not how much of your repo fits. The figure is written two ways: the Gemini 3 family is documented as 1M / 64k output, the 2.5 family as the exact 1,048,576 input / 65,536 output (2^20). The literal limit appears in errors — an oversized 2.5 prompt fails with input token count … exceeds the maximum number of tokens allowed (1048576).

Auto routing is on by default, managed by the ModelAvailabilityService. It classifies each prompt and routes accordingly: simple → gemini-2.5-flash; complex → gemini-3-pro-preview if Gemini 3 is enabled, else gemini-2.5-pro. Picking Pro in /model instead biases toward the most capable available model (Gemini 3 Pro when enabled). The CLI's own internal calls — prompt completion, classification — use gemini-2.5-flash-lite and silently fall back through gemini-2.5-flash to gemini-2.5-pro without prompting or changing your configured model.

Your sign-in method, not the model menu, is the biggest lever on what you can run. The CLI supports three: Sign in with Google (OAuth), a Gemini API key (export GEMINI_API_KEY="...", key from aistudio.google.com/app/apikey), and Vertex AI (needs GOOGLE_CLOUD_PROJECT + the Vertex AI API enabled). The Google-account path gives the Code Assist Individual free tier — 1,000 requests/day across the full family, the only free way to reach the Pro models. The free API-key path is Flash-only at 250/day. Pay-as-you-go on an API key or Vertex removes the daily cap and bills per token.