# I think heavy AI users may be wasting more capacity on routing than on prompting — retained record

> Retained reddit record from reddit:artificial.

Canonical: https://highsignal.app/data/records/8c0d538dcb43ea18

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:artificial)  published 2026-09-25 11:19 ingested 2026-09-26 02:35

[original source ↗](https://www.reddit.com/r/artificial/comments/1wptoj3/i_think_heavy_ai_users_may_be_wasting_more/)

## retained text



```
I’ve been noticing something counterintuitive while running a multi-agent system for real work.

When I use one long Codex/ChatGPT session manually, I usually make one coarse decision at the start:

Which model? Which reasoning level?

Then that same configuration handles planning, research, trivial checks, hard reasoning, revisions and often validation.

That feels increasingly inefficient.

In the system I’m experimenting with, the work is decomposed instead.

Some steps are deterministic and use no model at all.

Some go to a cheaper/faster model.

Hard or ambiguous steps get stronger reasoning.

Validation is separate.

If the first choice fails, the system can escalate instead of paying the maximum cost from the beginning.

The surprising result is that several smaller model runs can appear to consume much less paid capacity than one large manual run.

I’m not claiming this is proven yet. Quota burn is affected by context size, tools, background/subagent activity and platform-side efficiency issues too, so “users just pick the wrong model” would be an oversimplification.

But it makes me wonder whether model selection should be treated as resource allocation rather than a chat preference.

In a company, you don’t assign every task to your most expensive senior expert. You route work to the least expensive level that can reliably handle it, then escalate exceptions.

Maybe AI systems should work the same way.

For people running agents at scale: are you actually measuring model/reasoning mix and capacity per verified result, or mostly choosing a model for the whole workflow?
```



document fetched 2026-09-26 00:19 · canonical https://reddit.com/r/artificial/comments/1wptoj3/i_think_heavy_ai_users_may_be_wasting_more
