GitHub Copilot adds prompt caching, smarter model selection

What happened

GitHub Copilot for VS Code now uses prompt caching to reuse model state across repeated requests instead of recomputing the same information, and tool search to load tool definitions on demand rather than sending every full tool schema into context on every turn. Auto model selection, already live in VS Code, github.com, and mobile, is expanding to Copilot CLI, GitHub App, and additional IDEs, with plans to make it the default for Copilot Free and Student plans.

Why it matters

As Copilot takes on longer agentic work — planning, editing, debugging, reviewing, and calling tools across extended sessions — these changes let the model focus more effort on the actual task rather than repeating context and processing unnecessary tool definitions. Copilot's Auto system routes to the model that fits the work: simpler tasks like quick explanations can use more efficient models, while complex multi-file changes get stronger reasoning when needed, without requiring developers to manually choose models each time.

What to watch

Auto was trained on conversations across 16 language families and stays within 4 points of English baseline routing accuracy across language groups. The router learns where stronger models add value by comparing responses from less capable and more capable models on quality dimensions, ensuring escalation to stronger reasoning happens only when the task actually benefits from it.