GitHub is reducing Copilot’s model costs by optimizing whole coding tasks instead of blindly shortening every tool response. Its tests found that aggressive compression could backfire: when useful details disappeared, agents reopened files or reran commands, adding turns, time, and tokens.
The shipped compressor therefore leaves source-like output from commands such as git diff and arbitrary scripts unchanged, reorganizes search results without dropping matches, and targets repetitive build, install, test, and progress logs. Agents retain a direct path to the original output, and GitHub monitors how often they need it as a signal that compression removed too much.
A simpler saving came from file reads. Copilot’s view tool was still prefixing every line with a number even though current editing tools no longer use those numbers. Removing the prefixes cut model-inference cost by roughly 5% in offline coding benchmarks. An online Copilot CLI experiment reduced average daily inference cost per user by about 3%, with no material decline in the tracked quality or satisfaction measures.
GitHub also shortened repeated task-tool instructions and now includes completed background results in notifications, avoiding an extra retrieval turn. The company cautions that prompt changes need behavioral tests: an early rewrite accidentally made independent agents run sequentially, and it was stopped before a smaller, tested instruction restored parallel work.