Gemini model behaviour

The Gemini API was not designed as a variant of anybody else’s. It has its own request body, its own names for things that exist elsewhere under different names, and a few concepts—grounding metadata, thinking budgets, explicit context caches—that have no equivalent in the chat-completions shape most people learned first. Nearly every surprise it produces is a consequence of that vocabulary rather than of the model: an empty answer is usually a finishReason you did not read, a dropped instruction is usually a system_instruction that was posted inside contents, and a truncated JSON object is usually an output ceiling doing exactly what it is documented to do.

These pages take the API on its own terms. Each one starts from the field as Google documents it—the literal key, the literal enum values, the literal place in the body—and works out to the behaviour you can observe. Where a figure is per-model or has changed between generations, it is named as such and pointed at the live reference rather than frozen into prose, because Gemini’s numbers move more than most.