The Million-Token Codebase

A larger input window can include more code and documentation in one request. Its value depends on the task, corpus, permissions, model behavior and cost. Measure the material with the intended tokenizer, preserve source versions and leave room for instructions, tools and outputs before assuming a repository fits.

Google’s Gemini 3.1 Pro Preview documentation, updated 18 August 2026, lists a 1,048,576-token input limit. Capacity is distinct from successful reasoning over all that material. RULER’s 2024 study evaluates long-context behavior in a specified cohort of models and synthetic tasks; it does not establish a common failure threshold for all 2026 models. This guide compares full selected context, retrieval and caching through task acceptance, source control, observed latency and total cost.