Cost per Task

Anyone operating an AI system pays for computing power. Providers usually bill this in very small units, for example per piece of text or per second of computing time. But these individual prices say little about what a real piece of work ultimately costs. This is exactly where cost per task comes in: you take a completed task as the yardstick, such as “fully answer a customer email” or “summarize a page of contract.” Then you add up everything that was needed for it – including the failed attempts and the revisions. The result is a number that can be directly compared with the cost of human labor.

Why individual prices are deceptive

A cheap model can be more expensive than an expensive one. That sounds contradictory, but it’s easy to explain. A weak model often needs three or four attempts before the result is usable. A strong model gets it right on the first try. If the individual price is five times higher, but the success rate is eight times better, the expensive model still wins.

On top of that comes hidden work that appears on none of the provider’s invoices. Someone has to check the results, correct them, or request them all over again. In many companies, this human time is the biggest cost block. Anyone who only looks at the price per piece of text misses it entirely.

For investors and companies, the metric is therefore the more honest yardstick for comparison. It shows whether using AI is economically worthwhile. If the cost per task falls below the wage for the same work, this changes entire industries. If it stays above, the technology remains an experiment.

What goes into the calculation

You start with the direct computing costs for one run. This includes everything the model reads and writes as text. Modern systems often think in multiple steps, check themselves, or call on tools like a search engine. Each of these steps costs extra. A single task can thus consist of dozens of requests.

Then the success rate comes into play. Suppose one run costs two cents and works in half of all cases. Then on average you need two runs, so four cents. You divide the cost by the probability of success. This simple division explains why reliability is economically so valuable.

Finally, you add the incidental costs. These include human oversight, data storage, and the effort of keeping the system running. A common mistake is to confuse cost per task with the price of a subscription. A subscription is a fixed monthly amount, whereas cost per task grows with the volume of work.

The figure in quarterly reports and product pricing

In news about AI companies, the metric often appears indirectly. When a provider announces that a certain task is now ten times cheaper, this is exactly what they mean. Such leaps arise from smaller models, more efficient computing methods, or cheaper chips. For investors, the trend in this figure is an important indicator of whether a business model is becoming viable.

The influence is also noticeable in software price lists. Some providers charge money per completed transaction instead of per user and month. A customer service tool then costs, for example, a fixed amount per resolved ticket. The provider itself thus bears the risk if its AI needs several attempts.

In everyday life, you encounter this principle when free offers suddenly get limits. Image generators only allow a few images per day, chatbots switch to a weaker model after a few requests. The reason is always the same: every task costs the provider real money, and at some point giving it away for free no longer pays off.