Researchers from Microsoft and the University of California, Berkeley have released a free tool that can cut the cost of using an artificial intelligence coding assistant by almost a third. Called TokenEconomizer, it sits alongside assistants that programmers already use rather than replacing them. In tests on 277 programming tasks, it reduced spending by between 18.4 percent and 30.9 percent, depending on the assistant and model combination, while work quality mostly stayed the same or improved slightly. For developers who run an assistant many times a day, even the lower saving could add up quickly.
How TokenEconomizer Works
AI coding assistants charge, in effect, for the amount of text they read and produce, measured in tokens. Every file opened, instruction read, search run or solution tried consumes tokens, which paid users pay for. The researchers found that assistants often waste a large share of those tokens, repeating searches they have already performed, loading tools they do not need, and retrying approaches that have already failed. Across a full working day, that waste can quietly inflate the bill.
TokenEconomizer tackles that waste in three ways. It remembers useful facts about a project from earlier sessions, such as the commands needed to test the code, so the assistant does not hunt for the same information again. It strips out tools that are irrelevant to the particular task, leaving the assistant with less material to process. And it notices when the assistant is stuck repeating the same failed attempt, pushing it to try a different approach or stop spending resources on a dead end. It works alongside existing assistants, so no model switch or new environment is needed.
Savings of Up to 31% in Testing
The savings held up across every combination tested. With Claude Code running the Sonnet 5 model, costs fell by 23.6 percent, while tasks completed correctly on the first attempt rose from 74.7 percent to 76.0 percent. The largest saving came with Claude Code running Opus 4.8, where costs dropped by 30.9 percent, although first-attempt success slipped slightly, from 74.0 percent to 73.0 percent. OpenAI’s Codex with GPT-5.6 Terra cost 24.7 percent less and improved from 73.8 percent to 74.7 percent, while Codex with GPT-5.6 Sol cost 26.1 percent less and improved from 80.1 percent to 81.7 percent. GitHub Copilot CLI with Sonnet 5 recorded the smallest saving, 18.4 percent, yet its success rate still rose from 75.0 percent to 77.6 percent.
That matters because a cheaper tool producing worse code would be a false economy. In four of the five combinations, first-attempt success actually rose slightly with TokenEconomizer in place. The researchers caution, though, that real savings will vary with the assistant, the project and how often the tool is used.
A Tool That Improves With Use
TokenEconomizer also gets better the longer it stays with a project. Early on it is still learning, and a few projects even cost slightly more at first. By the sixth task, however, every one of the six projects studied had become cheaper to work on, as the stored knowledge began to pay off. Separately, removing unused tools saved 13.2 to 17.7 percent, remembering project information saved 11.8 to 14.3 percent, and stopping repeated failures saved 5.4 to 7.1 percent.
How It Compares With Other Cost-Cutting Tools
TokenEconomizer is not the first tool aiming to make AI coding assistants more economical, and the researchers compared it with RTK, Serena and Aider RepoMap. RTK reduced costs by 8.9 percent at best but increased them by up to 4.5 percent in other cases, raising costs in 12 out of 30 combinations. Aider RepoMap cut costs by up to 6.1 percent in some cases but raised them by up to 6.5 percent in others, with higher bills in 15 out of 30 combinations. Serena was not in the main cost tests. By contrast, TokenEconomizer reduced costs in all 30 combinations tested, with no case in which spending rose. The researchers say this consistency comes from adapting to each project rather than applying fixed rules everywhere.
Free to Use With Popular Coding Assistants
The practical requirements are modest. TokenEconomizer supports Claude Code, OpenAI Codex, GitHub Copilot CLI and Pi, runs on Python 3.10 or newer, and needs no dedicated graphics card. It needs no API key of its own, though paid access to the assistant itself is still needed. What it learns is stored locally, not sent elsewhere, reassuring developers working on private code. It can be downloaded from GitHub under an open-source licence allowing use and adaptation.
The broader significance is clear. Coding assistants are now standard equipment for many developers, but running costs concern freelancers, students and smaller companies most. Making existing assistants more efficient, rather than demanding pricier models, could keep those benefits within reach. TokenEconomizer will not remove AI coding costs entirely, and 30.9 percent is the best test result, not a promise for every user. Yet as a free, open and locally run way to trim waste from a daily workflow, it is welcome, and independent use will show whether its laboratory savings hold up in practice.
