What a token really costs you

When you price out an AI feature, the first number you find is the price per token. Every provider lists it, usually as dollars per million tokens. It is easy to compare, so it feels like the number that decides your bill.

It does not. The price per token is a rate, not a total. What you actually pay is that rate times the number of tokens your task uses, and the token count is the part nobody puts on the pricing page. Two models can charge the same per token and still cost very different amounts for the same job, because one uses far more tokens than the other.

The same job can use very different amounts of tokens

Here is a real example. We gave three Grok models the same job, pulling a few action items out of a short meeting note. Their input prices per million tokens are close:

• Grok 4.3: $1.25
• Grok 4.6: $2.00
• Grok 4.7: $2.00

From those prices you would expect the priciest model to cost about 60% more than the cheapest. The real gap was much wider, because the models did not use the same number of tokens to answer.

Same task. Grok 4.3 answered in about 400 tokens; Grok 4.7 took over 1,400.

Grok 4.3 answered in about 400 tokens. Grok 4.7 took over 1,400 for the same note. Nothing about the task changed. The newer model simply works through more text to reach its answer. You pay per token, so more tokens means more cost, on top of the higher rate. The result was that Grok 4.7 cost roughly five times what Grok 4.3 did for the same job, not the 60% the price list implied.

So the price per token told you part of the story. The token count told you the rest.

Two more things the rate leaves out

Token count is the big one, but a couple of others move the real cost too, and they are worth knowing before you commit to a model.

• Output usually costs more than input. Most providers charge more for the words a model writes than for the words you send it. A model that answers briefly can be cheaper than a wordier one even at the same headline rate.
• Some work is billed at a discount. Jobs that do not need an instant answer can often run in a cheaper batch mode. Text you send on every call, like a fixed set of instructions, can be cached and billed at a fraction of the normal rate after the first time. Both lower the real cost below the sticker.

None of this shows up when you compare price-per-token numbers side by side.

An easy job: the cheap model is fine

Not every task needs a strong model. We gave five models from four providers the meeting-note job above and scored each answer against a correct version. Four of the five landed at the top, within a few points of each other. Quality was not the thing that separated them. Cost was.

On a job like that, the question answers itself: pick a model that clears your bar and costs the least to run. The expensive models earned nothing, because the job was not hard enough to need them. This is most day-to-day work, and it is where inexpensive models shine.

A harder job: the cheapest model is the wrong choice

Now a task with a single right answer. Same five models, but this time they had to work out shipping dates under a few rules, including a holiday that pushes one order to the next working day.

Cost per run on the harder task. Green cleared the quality bar of 90; red fell short. The cheapest model failed; a mid-priced one was the cheapest that passed.

Here is what happened, lined up against each model’s price per token:

• GPT-4.1 Nano, the cheapest model per token, got the holiday wrong and scored 75. It failed the bar. Cheapest per token, but useless for this job.
• GPT-4o Mini and Claude Haiku 4.5 also scored 75. They failed too.
• DeepSeek V4 Flash got every date right, scored 100, and cost about $0.00016 a run.
• GPT-4o, the most expensive model per token, also scored 100, but cost about $0.0011 a run, roughly seven times DeepSeek.

So the cheapest-per-token model could not do the job, and the most expensive one did the job but cost seven times more than necessary. The right answer was the model in the middle. You would never have picked it by reading the price list, because the price list does not know your task.

The right answer does not stay the right answer

Even once you find it, it does not stay fixed. Providers change these prices often, and they ship new models constantly. A new flagship is usually expensive and aimed at hard problems, which is exactly the work your cheap day-to-day tasks do not have. Last month’s mid-priced winner might be undercut by a cheaper model that just got better, or your current pick might quietly get more expensive.

Illustrative. Both models hold the same quality, but the cheapest good-enough choice flips as one creeps up and the other drops.

You can sketch the stakes with simple arithmetic. Take the two models that scored the same on the shipping-date task, one at $0.0011 a run and one at $0.00016, and project the monthly bill as you scale up.

Same answer, very different bill. The gap only grows as you scale.

A quick projection like that is worth doing before you commit, and worth redoing when prices or models change.

The way to keep this current is to run your own task across a few models on a schedule and watch the cost-per-answer next to the quality score, so you always know the cheapest model that still clears your bar. That is what the AI API pricing pages and what AI API pricing doesn’t tell you go into, provider by provider.