My AI did the same task twice. Second time cost 75% less and did 5x more work. Here's what happened.

A few weeks ago I ran two tasks back to back on the same platform.

Task 1: collect 20 posts from X about a topic. It took 65 credits.

Task 2, maybe 10 minutes later: same kind of job, different keyword, but this time I asked for 100 posts. It cost 16 credits.

Let me say that again. 5x more output. 75% cheaper. On the second run.

I didn't do anything differently. I didn't optimize a prompt. I didn't change a setting. The only thing that changed was that the agent had already done it once.

Here's what actually happened under the hood.

The first time, the agent had never worked with X search results before. So it explored - figured out the page structure, learned how posts load, worked out how to extract them reliably. That exploration cost tokens. Then it completed the task and saved what it learned as a reusable workflow.

The second time, it read the workflow it had just built and went straight to execution. No exploration. No re-planning. Just work.

Same input from me (a one-line task description). Completely different cost profile.

We ran a third task a bit later - collect posts plus full author profiles for each one, something the agent hadn't done before. That cost 123 credits because it had to explore author profile pages for the first time and build a new workflow.

Then a fourth task: same thing, but 5x more data. Cost: 32 credits.

So the arc looks like this:

Task 1 - 20 posts, new platform: 65 credits
Task 2 - 100 posts, same platform: 16 credits
Task 3 - posts + profiles, new capability: 123 credits
Task 4 - posts + profiles, 5x data: 32 credits

Every time the agent hits something new, it pays the exploration cost once and saves the result. Every time it returns to known territory, it collects the dividend.

The thing that surprised me most: I didn't have to ask it to do any of this. It just did it. Saved the workflows automatically, reused them automatically. The compounding happened without me managing it.

We call this ROTI - return on token investment. The idea is that the first run of any task is an investment, not just a cost. You're paying for exploration that you'll never have to pay for again.

Most AI tools treat every task as a fresh problem. You pay full price every time. The tool never builds on what it already knows.

I think this is actually the core thing wrong with how most people think about AI costs. They look at cost per task. The right metric is cost per task over time, as the agent accumulates knowledge.

Anyway - we're building this into a product called AllyHub (allyhub.ai). Still in closed beta. If you're curious about the compounding angle or want to try it, drop a comment or DM me.

Has anyone else been tracking how their AI costs change over time as they use the same tools repeatedly? I'm curious whether others are seeing similar patterns or if our numbers are unusual.