RTK (Rust Token Killer) is a popular tool designed to filter and compress terminal output before it is processed by AI coding agents, aiming to reduce token consumption and thus lower costs. With over 79,000 stars on GitHub, RTK has attracted attention for claims of cutting token usage by up to 90% in some cases. However, recent independent testing challenges the extent of these savings.

A detailed benchmark conducted using Terminal-Bench 2.1 tested RTK’s impact on two AI coding platforms: Claude Code with Fable 5.0 and OpenCode with DeepSeek V4 Pro. Each task was run multiple times with and without RTK to compare token costs and pass rates. The results showed only a 5% cost reduction for Fable and a 5% increase for DeepSeek when using RTK. Pass rates slightly declined by 1-2% with RTK.

When factoring in all attempts, including failures, Fable’s cost was 3% lower with RTK, while DeepSeek’s was 7% higher. On a per-task basis, DeepSeek’s costs increased by an average of 17% with RTK, and Fable showed no significant difference. Notably, most of Fable’s savings came from a single task, with minimal impact elsewhere.

RTK compresses terminal output by rewriting shell commands to produce terser results, such as removing file owner and date details while retaining essential information like file permissions and sizes. Despite reporting large token savings based on output reduction, these figures do not directly equate to lower billing costs. RTK’s internal metric counts filtered output bytes rather than actual billed tokens, and it does not account for additional AI agent interactions that may increase overall token usage.

The benchmark also highlighted cases where RTK caused inefficiencies. For example, a DeepSeek task entered a retry loop due to an unsupported command rewrite, significantly increasing costs. Although this issue was fixed in a later RTK version, it illustrates how output compression can sometimes backfire.

Further analysis showed that terminal output constituted a relatively small portion of total input tokens—about 11% for Fable and 40% for DeepSeek. RTK was applied to a subset of shell commands, while other file operations bypassed it. In some cases, RTK reduced terminal output characters but led to more AI agent turns, which increased overall token consumption.

The findings suggest that modern AI coding models already employ efficient terminal output handling techniques, such as limiting output with commands like head or tail. RTK’s compression may have been more beneficial with older models but currently offers only niche optimization rather than broad cost savings.

Given these results, the benchmark does not recommend RTK as a general-purpose tool for reducing AI coding costs. The study underscores the importance of evaluating token savings in the context of total AI interaction costs rather than relying solely on output compression metrics.

This analysis was conducted using RTK version 0.45.0, Claude Code 2.1.220, OpenCode 1.18.25, and Harbor 0.20. Further research and benchmark data are available upon request.