July 15, 2026
Does RTK really save 60–94% of tokens?
Users reported dramatic reductions from filtering verbose tool output. Critics argue the headline measures removed text, not necessarily billable tokens that would otherwise reach the model.
What happened
Users reported 94.1% output reduction across 86,903 commands and 86.8% across 5,501 commands. RTK filters verbose tool output before it enters agent context.
What the evidence shows
The underlying method is supported by broader work on tool-output filtering and context compaction. Independent systems report meaningful reductions under narrower evaluations.
What appears true
RTK is likely a material optimization, especially for repetitive shell output. But smaller tool output is not automatically equivalent to the same reduction in context usage or billing.
What remains uncertain
Built-in truncation may already exclude some of the measured text. A credible test needs identical tasks, provider-reported token totals, success rate, repair cost and elapsed time.