Microsoft and UC Berkeley release TokenEconomizer, cutting coding-agent costs 18–31%; GitHub repo hits 5,400 stars

Microsoft Research and UC Berkeley researchers have released TokenEconomizer, an open-source plugin that cuts the token cost of coding agents such as Claude Code and Codex by 18.4–30.9% on average.

According to the GitHub repo, on SWE-Bench tasks mean cost fell for all five agent–model combinations and was lower than default in all 30 repository-by-combination cells. Pass@1 rose in four of the five. Claude Code with Sonnet 5 saw cost drop 23.6% with Pass@1 up 1.3 points; Claude Code with Opus 4.8 saw cost drop 30.9% with Pass@1 down 1.0 point. These are the authors' own measurements.

It works in three ways: "micro-skills" that tell the agent repo facts (for example, using `python -m pytest` instead of `pytest`) once three sessions confirm them; "tool pruning" that drops unused tool definitions from context; and "failure-loop intervention" that nudges the agent to try another approach after the same error three times. It supports Claude Code, Codex, GitHub Copilot CLI and Pi under Apache 2.0.

The repo was created Oct. 13 and had 5,394 stars today. Developer channel Better Stack's Nov. 8 install-and-test video has about 58,000 views. A top comment pushed back: "Claude Code is very good at memorizing such things. No need for a special harness" (18 likes).