GitHub Copilot is improving token efficiency through two main mechanisms: prompt caching and deferred tool loading in VS Code, and Auto model selection that routes tasks to the most appropriate model. Prompt caching reuses model state for repeated prompt prefixes, while tool search loads tool definitions on demand rather than sending all schemas upfront. The Auto feature uses HyDRA, a routing model that considers task complexity, reasoning depth, and real-time model health to pick the best-fit model without requiring manual selection. Auto is cache-aware, avoiding mid-conversation model switches that would break cached prefixes. It supports 16 language families with routing accuracy within four points of the English baseline. Auto is expanding to Copilot CLI, GitHub App, and additional IDEs, and will become the only model option for Free and Student plans. Practical tips include keeping context focused, avoiding mid-session model changes, planning before using parallel agents, and limiting enabled tools to what's needed.
