AI NewsProductivityAnnouncement
TokenEconomizer cuts AI coding agent cost 18 to 31 percent on SWE-Bench
Microsoft Research released a free plugin that spends 18 to 31 percent less per task on SWE-Bench.
Image: Microsoft Research
Why it mattersTeams running Claude Code, Codex or Copilot CLI on big repositories can cut the bill. They do not need to change the agent or the model.
Running a coding agent on a large codebase gets expensive fast. Teams see the bill. They do not see why. Microsoft Research and UC Berkeley put a number on the waste on 16 November 2026. On SWE-Bench, 30 percent of an agent's tool calls repeated something it had already done in the same repository. And 16 of the 23 tools shown to it in a turn were never called.
TokenEconomizer is the free plugin they released at the same time. It sits beside an existing agent and reads the traces of past tasks. It then acts on them in three ways. It writes one-line notes about the repository, once a pattern has shown up enough times. It hides the tools the recent tasks never used, so the prompt is smaller next time. And a judge model watches for a loop and ends the agent's turn when it is stuck.
The numbers, with the method
The paper tested five agent and model pairs on 277 bug-fix tasks from six SWE-Bench repositories. On that set, TokenEconomizer cut mean cost per task by 18.4 to 30.9 percent. Failed tasks had been 17 percent of runs and 32 percent of spend. That is where the loop check pays off. Pass@1 went up in four of the five pairs. It held steady in the fifth. So the savings do not come from giving up.
The authors say the plugin starts with no history. Every repository then becomes cheaper from the sixth task onward. Total cost fell below the baseline in every repository they tested. On a second test, Web-Bench, savings were 16.0 to 21.2 percent. On a security auditing test, 6.4 to 19.6 percent.
What runs it
Microsoft says TokenEconomizer works with Claude Code, Codex, GitHub Copilot CLI and Pi. The project page names three existing plugins as its comparison baselines. They are RTK, Serena and RepoMap. So the savings are measured against other cost-cutting tools, not against a bare agent. The code is on GitHub under Apache 2.0 at microsoft/TokenEconomizer. It had 7,100 stars at the time of release.
The authors are Aaron J. Li, Bin Yu and Ion Stoica at UC Berkeley. Microsoft Research adds Weijia Xu, Chandan Singh, Jeevana Priya Inala, Michel Galley and Jianfeng Gao. Microsoft Security adds Zixiao Chen, Charlotte Siska and Acar Tamersoy.
The design is worth more than the number. Tool hiding and short notes are cheap. But they need a trace of real work to be useful. So the savings grow with each task the agent finishes. A team that keeps session logs has most of what the plugin reads. Teams that throw traces away will see the smaller first-run number. Keep them.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

