Microsoft and Berkeley Release TokenEconomizer: Coding Agents Run 18–31% Cheaper

The bill for coding agents is growing faster than it seems. A single run of Claude Code or Codex on a non-trivial task involves dozens of steps, and at each one the model re-reads the accumulated context. That is why any promise to cut costs without changing the model gets the attention of teams for whom tokens have already become a separate budget line.

That is exactly the promise made by Microsoft and UC Berkeley: they have open-sourced TokenEconomizer, a tool for making agents cheaper to run.

What is claimed

According to a post on X, the team ran the tool on 277 SWE-Bench tasks. Five agent-and-model pairings were considered. The average cost of execution fell by 18.4–30.9% depending on the pairing. The top figure, 30.9%, is in the headline, but it is the best case, not a guarantee for every stack.

The key phrase is "without touching the model." The optimization works from the outside: no fine-tuning, no swapping to a cheaper model, and no access to weights is required. For those who use closed models via API, this is the only realistic lever: you do not control the price per token, but you do partly control the number of tokens.

The post's author writes that the whole technique fits in a single repository and three steps. The post does not say what those steps are; it only links to the repository. So it is more honest not to guess at the mechanics, but to read the code.

Why it matters in practice

In agentic scenarios, cost is not made up of a single request but of a chain. Context, tool output, retries: all of it is paid for again and again. Saving 20–30% on such a stream is not cosmetic. For a team spending several thousand dollars a month on agents, the difference is noticeable, and for a startup with a coding agent at the core of its product, it can shift the unit economics.

There is a second aspect. The work from Microsoft and Berkeley shows that agent scaffolding still has a lot of inefficiency. While everyone argues about which model is best, part of the money leaks out at the orchestration level. That is a signal to product teams: it is worth optimizing not only the choice of model but also how the agent handles context.

What to check

First, the metrics. The post talks about cost but says nothing about the percentage of tasks solved. If the savings come at the price of lower quality, there is no gain. You need to see whether the success rate stayed at its previous level.

Second, transfer to reality. SWE-Bench consists of neatly packaged tasks from popular open-source repositories. Your private monorepo with a messy commit history is built differently, and 18–31% there could turn into 10% or, conversely, 35%.

Third, the source. The post got few views, and the figures are relayed secondhand. Before building the savings into a budget, open the repository and, if there is one, the paper, and check the experimental conditions: agent versions, prices, and how cost was calculated.

What to do

The practical plan is simple. Clone the repository, run it on 20–30 of your typical tasks with the tool on and off, and compare cost and the share of successful solutions. If the savings hold up at even half of the claimed level, adoption will pay for itself within days. If not, you will lose a couple of hours and get your own data instead of someone else's percentages.