---
description: Small single tasks and large repeated workflows can create very different amounts of work.
url: https://www.schmalzl.tirol/en/blog/2026-04-13-ki-verbrauch-umfang-parallelitaet
lang: en
date: 2026-04-13
deutsch: /blog/2026-04-13-ki-verbrauch-umfang-parallelitaet.md
---

# Understanding AI usage without pricing promises

_13 April 2026 · Archive date_

A short answer and a longer workflow can come from the same tool. Even so, it makes sense to look at their scope separately. Revising one paragraph once is a different task from reading a hundred files, making changes and checking the results several times.

If you want to understand usage, look first at the work that is actually triggered. A single price figure does not answer that question.

## An invented comparison

Take ten notes you wrote yourself. Variant A sends all ten to a model again for every small change. Variant B provides only the affected note and the rules it needs. The second task is more narrowly defined; whether the specific service charges less for it has to be checked against its billing.

Another variant runs several tasks at the same time. That can shorten a wait, but it still creates several jobs. If the tasks then check each other again, more steps are added.

For a traceable experiment, I would record the number of calls, the context provided and the repetitions needed. It also matters whether the result was usable at all. Low usage per answer helps little if five attempts are needed.

## Separating observations from pricing questions

Current prices, quotas and billing rules belong in each provider's documentation. Personal calculations from a group discussion are not a general basis for pricing.

You can still examine your own way of working: define tasks more narrowly, spot unnecessary repetition and set an end point. This makes it clear what work a tool really does, before a short product announcement is turned into a supposed savings comparison.

## Original finds

These links were the starting point for this post. The assessment in the text is mine; the linked posts are not my own tests.

- [Looking at Claude usage with more targeted tasks](<https://www.linkedin.com/posts/theanishajain_you-dont-need-the-100-claude-plan-instead-share-7449332534385192960-CQg5>)
- [AI costs: usage, limits and your own way of working](<https://www.instagram.com/reel/DXcootSBDYl/>)
- [AI costs: usage, limits and your own way of working](<https://www.linkedin.com/posts/dushyant-khosla_github-said-my-copilot-pro-plan-will-move-share-7455220189102497792-c_lv>)
- [AI costs: usage, limits and your own way of working](<https://www.linkedin.com/posts/timospringer_github-copilot-ist-auf-usage-based-pricing-share-7454771842898374656-mvqq>)
- [AI costs: usage, limits and your own way of working](<https://www.facebook.com/share/p/1BXncfzekX/>)
- [AI costs: usage, limits and your own way of working](<https://www.facebook.com/share/p/1Dfrdy7mbL/>)
- [AI costs: usage, limits and your own way of working](<https://www.instagram.com/reel/DYSKqaqO7SF/>)
- [AI costs: usage, limits and your own way of working](<https://www.linkedin.com/posts/philippdeutscher_die-tokenpreise-steigen-h%C3%B6re-ich-leider-share-7465639836876574720-wqxm/>)
- [Keeping an eye on AI usage](<https://github.com/robinebers/openusage>)
- [Public complaint about the costs of Claude and OpenClaw; to be read as a personal account.](<https://www.facebook.com/share/p/1HpAK4mCU4/>)
- [Example calculation for usage-based Copilot billing; the figures apply to the case described.](<https://www.linkedin.com/posts/marc-robin-zeeb_die-all-you-can-eat-illusion-im-ai-coding-share-7465372872090460160-g2-I/>)
- [Personal account of price rises for AI graphics cards; no current market prices confirmed.](<https://www.linkedin.com/posts/chriswojzechowski_offene-ki-modelle-sind-gratis-geworden-die-share-7468658602098978816-4oBI/>)
- [TokenTamer as a proxy for optimising the token usage of coding assistants.](<https://www.facebook.com/share/p/1Cyioch6SA/>)

> **About the date:** This post was written on 11 October 2026. The archive date matches the earliest related find in the collection, not a publication at that time. Technical assessments and sources added later refer to the time of writing.