Study finds AI agents waste 18% of tokens on unnecessary tasks
A new US academic study quantifies how AI agents waste tokens and money during agentic workflows. Researchers tracked token usage across six models—Kimi-K3, Gemini 3.1 Pro, GPT-5.6 Terra, GPT-6 Astra, Claude Sonnet 5, and Claude Fable 5.1—using a ‘shadow’ methodology to measure waste in closed-source systems. They found that 18% of output tokens were wasted, with 45% of tasks seeing one agent…
Key points
- Researchers found AI agents waste **18% of output tokens** on unnecessary tasks, with **45% of tasks** seeing one agent burn more tokens than needed
- Removing **41–46% of redundant reasoning** had little impact on performance, suggesting major cost savings are possible without sacrificing results
- Waste patterns include **restating info, duplicate tool calls, padding text, and unproductive detours**, with harness flaws like wrong Python environments causing retries
Model pages: GPT-5.6 Terra → · GPT-6 Astra → · Claude Fable 5.1 →
The story so far
4 episodes →- Study finds AI agents waste 18% of tokens on unnecessary tasksthis story
The AI Detours That You Shouldn’t Be Paying For
Unite.AI · 6 November 2026
Loading the full article…
This text was published by Unite.AI and written by Martin Anderson. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Perplexity launches Teammates agents for Perplexity Max users · 3 src
- Zapier Sidekick automates deal reports, Slack channels and more for users · 1 src
- Zapier’s Sidekick shifts AI approvals from tools to protected resources · 1 src
- OpenAI, Meta and xAI launch AI agents for long-term task delegation · 8 src
- Amazon Quick defines four agentic primitives based on When, What, Who, and How · 1 src
Comments
via GitHub Discussions