Microsoft, UC Berkeley release TokenEconomizer to cut AI coding agent costs

Researchers at Microsoft and UC Berkeley have released TokenEconomizer, an open-source plugin that reduces the cost of running AI coding agents by 18.4–30.9% while keeping task success rates steady, according to aitimes.com. The tool was published as open source on GitHub.

The plugin operates alongside major coding agents, including Claude Code, Codex, GitHub Copilot CLI and Pai, without changing model weights or internal structure. It requires no separate GPU.

In an evaluation covering 277 tasks across six repositories of the SWE-Bench benchmark with five agent-model combinations, costs fell 18.4–30.9% compared with existing execution methods. Pass@1 success rose in four of the five combinations and dropped by 1 percentage point in the remaining one. Cumulative costs dropped below the baseline from the sixth task onward in every evaluated repository. On the Web-Bench web development benchmark, savings of 16.0–21.2% were confirmed.

Where the savings come from

TokenEconomizer offers three features: Micro-skills, Tool Pruning and Failure-loop Intervention. In a Claude Code experiment, about 30% of tool calls made after five or more prior tasks repeated earlier actions, and 16 of the 23 tools loaded on each run were never used.

Failed runs carried a disproportionate share of spending: ten failed runs, amounting to 17% of analyzed tasks, consumed 32% of total cost. The work was also published on the arXiv preprint archive.