News · 2026-09-11

RTK claims up to 90% token savings for coding agents. A 1,740-run cost test found about 5%.

A cost test published on 11 September 2026 by the developer-tools company Quesma found that RTK, a widely used open-source tool that shrinks terminal output before an AI coding agent reads it, cut one agent's total bill by about 5% and raised another's by about 5%. RTK advertises that it "reduces LLM token consumption by 60-90% on common dev commands," and its own built-in counter reported an 89% reduction during runs that ended up costing more.

Key facts

• About 5%: RTK's effect on total cost, in opposite directions for two setups, across 1,740 attempts and more than $1,500 of spending.
• Setups: Claude Code with Fable 5.0, and OpenCode with DeepSeek V4 Pro; every task ran five times with RTK and five times without.
• Almost all of the saving in the Fable setup came from a single task, and with DeepSeek, RTK led to 18% more agent turns.
• Primary source: Quesma's write-up by Bartosz Kotrys and Jacek Migdal.

What RTK does

Coding agents work by running shell commands (listing files, running tests, reading logs) and then reading the output back as input. Long logs cost money, so RTK, short for Rust Token Killer, sits in between and trims the output first. Its own documentation is careful: it "cuts up to 90% of the bash output your agent reads" and "it is not the same as cutting your bill by 90%." Social media has been less careful, and the tool's popularity rests largely on the savings story.

What Quesma measured

Quesma used Terminal-Bench 2.1, a set of realistic tasks an agent must complete by working in a command-line terminal, and ran each task repeatedly with and without RTK in two very different setups. The result was close to a wash. RTK saved roughly 5% on the Fable setup's total bill, but almost all of that came from one task; across the rest, the saving was under one percent. On the DeepSeek setup it made things about 5% more expensive.

Quesma gives three reasons the savings evaporate. Terminal output is a small share of what a frontier agent reads: most of its input is instructions, tool definitions, conversation history and code. Much of that input is served from a cache at a steep discount (see prompt caching). And trimmed output can hide something the agent needed, so it spends extra turns getting it back. Shrinking the receipts in your grocery bag does not lower the grocery bill, and if you shrink them until they are unreadable you may have to go back to the shop.

The counter that misleads

The most useful finding is about measurement. "rtk gain counts removed output, not money saved, and it can make a more expensive attempt look optimized," the authors write. One command, head -1, was credited with 120.5 million tokens saved each time it ran. A separate bug sent one attempt into a loop of 339 errors that cost about nine times as much as the matching run without RTK; Quesma says it is fixed in version 0.46.0. Their conclusion: "We do not recommend RTK as a generic cost-saving tool."

Why it matters

This is a familiar trap in AI engineering: a tool reports a number that is easy to measure (tokens removed) rather than the number you care about (dollars spent), and the gap goes unnoticed because nobody runs the controlled comparison. The broader point about where agent costs really come from is covered in our lesson on inference cost and token economics. A separate benchmark published this week, Nine coding harnesses vs. your laptop, points at the same place from another angle: one popular harness sends a system prompt of about 18,000 tokens before the agent has read a single log line. The harness, not the terminal, is where most of the input comes from.

The honest caveats: this is one benchmark, two setups and five runs per task, and workflows that generate enormous logs, or models without cheap caching, might see more benefit. Both parties have incentives. Quesma sells tools for understanding what coding agents do and says it plans to test a competing product, Headroom, next. RTK's users have a stake in the savings being real. The Hacker News discussion includes developers reporting both outcomes.

Primary source, verified: read the paper →

Ground Truth. (2026, September 11). RTK claims up to 90% token savings for coding agents. A 1,740-run cost test found about 5%.. Ground Truth. https://groundtruth.day/news/rtk-claims-up-to-90-percent-token-savings-a-cost-test-found-about-5.html

@misc{groundtruth:rtk-claims-up-to-90-percent-token-savings-a-cost-test-found-about-5,
  title  = {RTK claims up to 90% token savings for coding agents. A 1,740-run cost test found about 5%.},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/rtk-claims-up-to-90-percent-token-savings-a-cost-test-found-about-5.html}
}

Topics: coding-agents · inference-cost · developer-tools · evaluation · benchmarks