← sources

What kind of input/output token ratios are you getting with long-running AI agents?

source: reddit (reddit:AI_Agents)published 2026-10-23 03:28ingested 2026-10-24 02:35

retained text

I've been experimenting with tooling designed specifically for long-running autonomous agents, with a focus on reducing context overhead, unnecessary repetition, and the amount of work agents spend just navigating their environment.

I've been looking at the token statistics lately, and I'm curious how they compare to what other people are getting.

Here's an example from an agent that ran for a little over 24 hours, so more of a shorter run:

    Total:      22,781,095
    Input:      16,706,366
    Cached:  1,090,578,176
    Output:      6,074,729
    Reasoning:   4,381,110

A couple of interesting ratios:

* **Input to output:** \~2.75:1 (excluding cached input)
* **Output relative to uncached input:** \~36.4%
* **Cached input:** \~98.5% of all reported input tokens, including cache reads
* **Reasoning:** \~72% of output tokens

I also resumed an agent that's been working on a compiler for a little over six days, and it was showing remarkably similar ratios.

Obviously, these numbers alone don't establish how productive an agent is. Output includes reasoning tokens, and generating more tokens isn't necessarily the same thing as accomplishing more work.

But I find the relationship between fresh input, cached context, and generated output particularly interesting. Especially when those ratios remain fairly consistent over extended periods of autonomous operation.

**What kind of ratios are other people seeing?**

Are your agents consuming substantially more input than they're generating in output? How much of that input is cached? And do those relationships change as your agents run longer?

I'm especially interested in people building custom tooling or agent environments, and whether you've found architectural decisions that noticeably change these ratios.

Just trying to establish some points of comparison for what efficient long-running agent execution actually looks like.

document fetched 2026-10-24 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xd35qc/what_kind_of_inputoutput_token_ratios_are_you