How much does it cost an AI model to do a business task?
The first edition is being prepared. It will appear here with its figures, its trend and the data to download.
Why we track AI cost per task
What does it cost to have an AI model do a real piece of business work, and how often does it get it right? Each month we run the same 20 small-business jobs through the main AI models, three times each, and publish what every answer cost and whether it passed a fixed check.
Price lists show cost per word; this shows cost per job done right, which is what a business pays for. Choosing a model? Read which AI model a small business should use and how to check the work AI does. We route every call through one provider pinned per model, as OpenRouter provider routing allows, so a change is the model, not the host.
Every edition of AI cost per task
Each edition is frozen when it is published; a later fix is a dated correction on that edition, never a silent change.
None published yet.
How AI cost per task is measured
Each month every model does the same 20 small-business jobs (task set 1 - pulling a total from an invoice, sorting an enquiry, tidying postcodes, removing personal details and 16 more), 3 times each, at temperature 0, one call at a time through one OpenRouter account with the model's own provider pinned. Each answer is marked pass or fail by a fixed check written for that job - no AI marks another AI. Cost is what OpenRouter billed for each call. Cost per passed task is the run's cost divided by the answers that passed, shown only when a model passed at least 3.
What it does not tell you
- "Passed" means an answer met our published check for that job - right number, right format, right facts. It does not judge tone, style or how helpful an answer feels.
- Twenty jobs is a small sample of what a business does; jobs that only a person can judge are left out on purpose.
- Prices and models change often; each edition is one month's run, and two models whose pass rates sit within each other's range across the 3 runs are not ranked against each other.
Method versions
- Version 1, 2026-10-26
Sources
How to cite AI cost per task
Quote a figure with the edition it came from - each edition has its own cite line. To cite the report as a whole, use this.
Eclipse Labs AI (2026). AI Cost per Task. https://eclipse-labs.ai/reports/ai-cost-per-task. Licensed under CC BY 4.0.
Measured by Eclipse Labs AI; costs are what OpenRouter billed us for each call.
Questions about AI cost per task
For everyday jobs like pulling a total from an invoice or sorting an enquiry, most AI models cost a fraction of a penny per answer; the newest edition on this page gives the exact figures, per model, as billed. The cost that matters is per PASSED task, because a cheap answer that is wrong costs you the time to fix it. See which AI model to use.
They are everyday small-business jobs with a right answer that a fixed check can mark: pulling a total from an invoice, working out VAT, sorting an enquiry, taking booking details from an email, tidying postcodes, adding up a timesheet, removing personal details before sharing, repairing a broken data file and twelve more. Each check is tested against a right and a wrong answer before any model is run.
No. It shows how often each model passed our checks on our 20 tasks, and what that cost, in one month. A model that does well here may do worse on your own jobs. Where two models' pass rates overlap across the three runs, we do not rank them. To test a model on your own work, read how to trial AI on one job.
Because AI models do not always give the same answer twice, even with the same settings. Running each task three times shows how steady a model is: the range on the table is its lowest and highest number of passes in a single run. A model whose range is wide is less predictable, which matters when a business relies on it every day.
Yes. Quote the figure with the edition it came from and a link; each edition has a cite-this line and a CSV of every model's results, free to reuse under CC BY 4.0. The tasks and their checks are published with the method, so anyone can repeat the run. See every report for the others we publish.