GitHub instrumented token usage across its production GitHub Agentic Workflows and began a systematic token-optimization effort in April 2026, according to a GitHub Blog post published May 7, 2026. The post reports that differing agent frameworks emitted usage logs in inconsistent formats and that an API proxy used by the workflows provided a reliable point to collect consumption data, enabling per-run accounting. GitHub describes applying logging, measurement, and targeted optimizations to reduce repeated and verbose model outputs. Community approaches such as the claude-token-efficient pattern and project-level rules files (the CLAUDE.md pattern) are cited as practical ways to cut output verbosity (drona23 repository). According to OpenAI, GPT-5.5 also uses significantly fewer tokens on comparable coding tasks, a model-level trend that can change cost calculations for agentic CI. Editorial analysis: For practitioners, combining measurement at the proxy layer with stricter output controls and choosing more token-efficient models offers the clearest path to cut recurring agentic CI costs.
