Microsoft unveils plugin to cut coding agent costs
BROBROK News · Based on reporting by DiarioBitcoin ·
Researchers from Microsoft Research and UC Berkeley introduced TokenEconomizer, an open-source plugin designed to lower the cost of coding agents by learning from prior sessions within each repository. In the reported tests, it reduced average costs by 18.4% to 30.9%, although one setup saw a small drop in accuracy.
Key points
- TokenEconomizer is an external, open-source plugin and does not change model weights.
- The evaluation covered 277 SWE-Bench tasks across five agent-model setups.
- It combines micro-skills, tool pruning, and failure-loop detection.
- Pass@1 improved in four configurations and fell by 1 point in one.
- The largest cost reduction was reported with Claude Code and Opus 4.8.
Why it matters
Coding agents can become expensive as they carry more context, tools, and retries through a workflow. A layer that trims those costs without retraining the model could affect how AI developer tools are built and priced.
Original headline: “Microsoft presenta TokenEconomizer para recortar hasta 30,9% el costo de agentes de código”
Read the original on DiarioBitcoin
Summary and analysis written by BROBROK News from the original report. Read the full story on the source website. Not investment advice.