RTK, or Rust Token Killer, is a tool aimed at reducing the token consumption of AI coding agents by filtering and compressing terminal output before it is read. It has gained popularity, with over 79,000 stars on GitHub, and claims to cut terminal output tokens by up to 90%. Despite this, recent benchmark testing suggests these savings do not directly translate into lower AI coding costs.
A detailed evaluation using Terminal-Bench 2.1 tested RTK with two AI coding platforms: Claude Code running Fable 5.0 and OpenCode using DeepSeek V4 Pro. Each task was executed multiple times with and without RTK to compare costs and pass rates.
The results showed a modest 5% cost reduction for Fable tasks when using RTK, but a 5% cost increase for DeepSeek tasks. Pass rates slightly decreased by 1-2% with RTK. When factoring in failed attempts, Fable’s cost savings narrowed to 3%, while DeepSeek’s costs rose by 7%. On a task-level basis, DeepSeek’s costs increased by 17% on average with RTK.
Most of Fable’s savings stemmed from a single task where RTK reduced the number of interaction turns by about half. Conversely, DeepSeek experienced more turns and higher costs on the same task with RTK. The tool’s own reported token savings, based on raw output reduction, did not align with actual cost reductions, as RTK’s compression sometimes led to additional agent interactions that increased overall token usage.
RTK operates by rewriting shell commands to produce shorter outputs but does not affect other tools that read files or perform searches, which can limit its impact. Additionally, AI agents often cache context between turns, reducing the cost of repeated terminal reads.
The study found that while RTK reduced terminal output characters by up to 9%, the total prompt tokens sometimes increased due to more frequent agent turns. This token inflation can negate the benefits of output compression.
Overall, the findings suggest that RTK is not a reliable general-purpose tool for reducing AI coding costs with current frontier models, which already employ efficient terminal output handling. RTK may have been more beneficial with older models but now serves as a niche optimization rather than a broad cost-saving solution.
These insights are important for developers and organizations seeking to optimize AI coding workflows and manage token expenses effectively. Understanding the nuanced impact of tools like RTK helps avoid overestimating their cost-saving potential.
The tests were conducted using RTK version 0.45.0, Claude Code 2.1.220, OpenCode 1.18.25, and Harbor 0.20. Further research and data are available upon request.