TokenEconomizer Cuts AI Coding Costs by 30.9%

Researchers from Microsoft Research and UC Berkeley have released TokenEconomizer, an open-source plugin built to cut the token bill run up by autonomous coding agents. It works as a cost controller that watches past sessions inside a repository and uses what it learns to trim waste in later ones, without touching model weights or the underlying agent code. The plugin plugs into Claude Code, Codex, GitHub Copilot CLI and Pi, and ships as a standard-library Python package with three switchable mechanisms, requiring only Python 3.10+ and no GPU or API key of its own.

The team tested it across 277 SWE-Bench tasks spanning five agent-model combinations, and found mean cost dropped between 18.4% and 30.9%, with savings showing up in all 30 repository-setting pairings they checked. The best result came on Claude Code paired with Opus 4.8, at 30.9% lower cost on its own and 33.2% when stacked with another technique called RTK. The weakest case was ReposVul audit work, where savings fell to 6.4% to 9.3%. To figure out where the money was going in the first place, the researchers profiled 60 SWE-Bench tasks across six repositories using Claude Code with Opus 4.8.