Microsoft Research and UC Berkeley released TokenEconomizer, a free plugin to cut coding-agent token costs
TokenEconomizer is an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security. It works with Claude Code, Codex, GitHub Copilot CLI and Pi, learning from past tasks in a repository to inject micro-skills, prune unused tools and break failure loops. It needs no API key of its own and does not touch model weights or agent code.
A video about it went from 34,638 to 59,867 views in 16 hours - about 8 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube. It has 717 likes and 57 comments, about 24 hours after it was published.
Why it matters: If coding-agent bills are a line item for you, this is a way to try lowering them without swapping models. The authors measured 18.4-30.9% lower mean cost on SWE-Bench tasks across five agent-model combinations, with Pass@1 up in 4 of 5. That is a benchmark result, so test it on your own repository.


