GitHub has rolled out two core updates to Copilot for VS Code focused on context handling and model routing, designed to cut redundant token consumption during long, agentic coding sessions. The changes target repeated context, redundant tool definitions, and unnecessary model compute that previously added fixed overhead to every multi-turn request, per the company’s official announcement.
In internal testing, the combined changes deliver up to a 40% reduction in token usage for extended agentic workflows. GitHub’s official Copilot for VS Code update announcement
GitHub Copilot context handling update cuts token waste with cached prefixes, on-demand tools
For extended Copilot sessions in VS Code, the previous harness sent full tool schemas, repository context, and full conversation history on every turn. This happened even when only a subset of tools were relevant to the current request. The new prompt caching feature uses cache-control breakpoints to identify static segments of the prompt prefix that can be reused across requests. These static segments include system instructions, repository metadata, and shared cross-turn context. This eliminates redundant recomputation of these segments for every turn.
Tool search, built with provider-specific logic for different LLM backends, lets the model load only the tool definitions required for a given request. The full available toolset includes four categories: MCP tools, terminal commands, file operations, and workspace search actions. For example, a multi-turn refactoring workflow that only needs file operation and workspace search tools previously received full schemas for all four tool categories, including unused MCP integrations and terminal command tools, on every turn.
GitHub notes these changes deliver the largest efficiency gains for agentic workflows that span multiple turns and use a broad toolset. Upfront schema costs previously added fixed token overhead to every request regardless of task needs. The 40% token reduction measured in internal testing applies specifically to these long, multi-turn agentic coding sessions using the combined context handling changes. GitHub’s official announcement of Copilot’s context handling improvements
Auto model routing selects the optimal LLM per task via HyDRA
Alongside the harness changes, GitHub has updated Copilot’s Auto model selection feature for VS Code to route individual tasks to the most appropriate LLM. It no longer defaults to a single model for all requests. The routing system uses a model called HyDRA that evaluates two core data sets for every request.
The first data set covers task complexity metrics: reasoning depth, code complexity, debugging difficulty, and tool orchestration needs. The second covers real-time model health data: availability, utilization, speed, error rates, and cost per token. In internal evaluations across a wide range of coding tasks, GitHub found no single model consistently outperformed others across all task types.
Smaller, more efficient models matched the output of larger, more expensive models for simple tasks like quick explanations or focused single-file edits. Stronger models only provided measurable quality gains for complex multi-file changes or deep debugging work.
For example, a request to explain a single utility function would be routed to a smaller, lower-cost model. A request to debug a cross-file memory leak would be routed to a larger model with stronger multi-step reasoning capabilities.
The company states the goal of the routing system is not to trade quality for cost, but to match model capability to task requirements. Unlike binary task classification systems that label tasks as simply “easy” or “hard,” HyDRA was trained on paired outputs from both low-capability and high-capability models, scored across quality dimensions.
This training lets the system learn exactly where model performance diverges across task types. It avoids over-escalating to expensive models for tasks where smaller models deliver identical results, while still routing complex work to models with sufficient reasoning capacity. GitHub’s internal evaluation of the HyDRA routing model
Routing logic preserves prompt cache efficiency across sessions
To avoid undermining the efficiency gains from prompt caching, Auto does not switch models mid-conversation unless a natural cache boundary is hit. These boundaries occur at the first turn of a session, or after a compaction event where Copilot summarizes older turns and resets the prompt prefix. For example, if a user starts a new 10-turn refactoring session, the routing system selects the optimal model for the first turn and uses that same model for all subsequent turns. This holds unless the conversation exceeds the token limit that triggers a compaction event.
GitHub notes that switching models between those cache boundary points would break the existing prompt cache. This would add more token cost than the routing change would save. The routing system also supports 16 language families beyond English. Evaluation accuracy stays within four percentage points of the English baseline across all supported groups. GitHub reports no statistically significant quality gap for non-English coding workflows as a result of this design. GitHub’s documentation of Copilot’s model routing cache preservation logic
