Researchers from UC Berkeley, Microsoft Research and Microsoft Security have released TokenEconomizer, an open-source plugin that cuts the token costs of coding agents, under Microsoft’s GitHub account. Their paper reports that it reduced average token cost by 18.4% to 30.9% across Claude Code, Codex and GitHub Copilot CLI paired with frontier models, “with comparable task success”.
The plugin also runs on Pi, the repository says, and is released under the Apache 2.0 licence. On the project page’s own profile, the 17% of tasks that fail take 32% of the spend, so a failing task costs about 2.3 times a succeeding one, by our arithmetic.
Short explainer · Microsoft’s free plugin, coding agent costs down up to 30.9%
A vertical news Short summarising this story. Figures come from the primary sources cited below; our arithmetic is labelled.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how an AI model and the harness around it split the work. It teaches the general idea only and does not discuss TokenEconomizer.
The key insight: TokenEconomizer’s authors report 18.4% to 30.9% lower mean token cost for coding agents, and their project page lists three kinds of recurring waste: rediscovering the same facts, paying for unused tools and getting stuck on failures. A failing task costs about 2.3 times a succeeding one, by our arithmetic.
The project page says every turn of a coding agent “re-bills the agent’s whole context”: its instructions, every tool description and the task so far. It lists three kinds of recurring waste from 60 profiled SWE-Bench runs of an unmodified agent.
First, “30% of tool calls repeat an earlier task’s action” once a repository has five or more tasks behind it. Second, “16 of the 23 tool descriptions billed on every turn are never called”, about 70%, by our arithmetic. Third, failing tasks take a share of spend nearly twice their share of tasks.
Left: share of tasks and share of spend for failing and succeeding tasks in the TokenEconomizer authors’ profile of 60 SWE-Bench runs of an unmodified agent. Right: the cost reductions the authors report with the plugin, by benchmark. The 2.3x ratio is FourWeekMBA arithmetic.
How the Plugin Cuts It
The paper describes TokenEconomizer as an external controller with three levers: which “operational micro-skills” to inject, which “persistently irrelevant tools” to prune, and when to intervene in “unproductive failure loops”.
The README says the plugin needs no API key of its own and never handles credentials. Its own model calls run through the user’s agent, and their cost is included in its reports. What it learns stays on the user’s machine, it says.
What the Tests Show
The README says the paper ran five agent-model combinations on the same 277 SWE-Bench tasks from six repositories. Mean cost was lower in “all 30” repository and combination cells, and Pass@1 rose in “4 of 5” combinations, the project page says.
Within each repository and combination, cost fell by 5% to 40%, the README says. On Web-Bench the savings were 16% to 21%, and on two security-auditing benchmarks where every task is in a new repository they were 6% to 20%.
The Business Lever: Cost per Task
The savings come from repeated work in the same codebase, which is how the paper says developers use these agents. The repository was created on 13 October and had 2,730 stars on GitHub when we read it on 31 October.
The Structural Read
As we read the paper, the savings come from the setting developers actually use: many tasks in the same codebase, where the same facts, tools and failure patterns recur.
As we read the README, cost is counted as billed, with cache pricing and the plugin’s own model calls included, so the reported cuts are net of its overhead.
As we read the README, the plugin works through what each agent already reads and through the agent’s own settings, without a change to the model.
The framework says “The new job of finance is to manage the ratio between the two — work produced per token burned.”
As we read it, a plugin that keeps task success comparable while cutting token cost raises work per token. The framework is the owner’s argument and does not discuss TokenEconomizer.
The results are the authors’ own measurements on benchmarks. The pages read do not report independent replication, savings on any company’s production workload, or which frontier models were paired with each agent in every run. We did not run the plugin.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
An open-source plugin from Microsoft Research and UC Berkeley researchers reports 18.4% to 30.9% lower mean token cost for Claude Code, Codex and GitHub Copilot CLI in the authors’ tests. Its three levers target context, tools and failure loops that recur in the same codebase.
A note on sourcing. All facts and figures come from the TokenEconomizer paper (arXiv 2610.22732), its GitHub repository and project page, read on 31 October 2026. Results are the authors’ own benchmark measurements; ratios are our arithmetic. We did not run the plugin or contact the authors.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
