Cheaper per Token Is Not Cheaper

The cheapest model on a price list is often the most expensive one in production.

Every model vendor publishes a price per million tokens, so that is the number everyone compares. It is the wrong number.

A model priced at a fifth of another is not cheaper if it needs five times the tokens, three retries, and a longer reasoning trace to reach the same answer. You paid less per token and more for the work.

The bill is not the price

Token pricing measures what you pay per unit of text. It says nothing about how many units the job takes.

Three things sit between the token price and the actual bill.

Reasoning traces. A model that thinks longer produces more tokens before it answers. Thinking tokens are real tokens. You pay for all of them.

Retries. A model that fails the first time and gets it on the third attempt cost three times its sticker price.

Length. A model that needs more words to reach the same answer costs more, even if every word is cheaper.

None of this appears on the price list. All of it appears on the bill.

The number that matters

What you care about is the cost to finish a task, not the cost of a token.

This is why we measure completed tasks, not token prices. We run real work through every model, score the output against a rubric, and record what it cost to get a correct answer. A model that needs three attempts to pass costs three attempts.

The result looks nothing like the price list. Some cheap models are cheap because they are efficient, not because they are worse. Some expensive models are expensive because they get it right once. The only way to tell is to measure the finished work.

Two kinds of tasks

Verifiable tasks have a right answer. Extracting a date from a contract. Classifying a support ticket. Summarizing a thread. Once a model gets these right, no newer model can get them more right. The cost should fall to the cheapest model that clears the bar, and stay there.

Graded tasks have no ceiling. Persuasive writing. Novel code. Complex reasoning. A better model produces a measurably better result, so these are worth spending on.

Most AI bills mix both, and most teams send both to the same model. That is the expensive part. Not the token price. The mismatch.

What this means for your bill

Stop optimizing the price per token. Optimize the price per completed task.

The cheap model that needs three tries is not cheap. The expensive model on a task any model can do is not worth it. Match each task to the model that finishes it correctly at the lowest cost, at your quality bar.

That is the whole job. It is a measurement problem, not a shopping problem.

Autark runs that measurement continuously, across a growing set of models, so each request goes to the model proven to finish that kind of work at the lowest real cost. You get a flat price and the finished result. The token math is our problem.

The bill is what you pay to get work done. Price it that way.