AI got cheaper again last week. For a service business, the model was never the expensive part.

On 22 September OpenAI released two new models, GPT-6 Sol and GPT-6 Luna, and priced them at roughly half of what the models they replace cost to run. The same day Anthropic released Claude Opus 5.5 and said typical workloads would cost around 40% less than on its previous top model. If you follow this space, you will have read a dozen versions of the same headline since: AI agents just got dramatically cheaper.

For a software company processing millions of requests, that is true and it matters. For a business with a few staff, a phone, a calendar and an inbox, it is true and it almost doesn't matter — because the price of the model was already the smallest number in the sum.

Do the arithmetic once

Take a realistic small job for an AI agent: read an incoming enquiry, look at the relevant part of your price list, and draft a reply. Generously, that is a few thousand words of input and a few hundred words of output.

At Luna's published rates — ten US cents per million tokens in, fifty cents per million out — that reply costs a small fraction of one cent. On Sol, the more capable of the two, it is in the region of a single cent. Even on the most expensive models on the market, a reply like that rarely reaches the price of a stamp.

Now put the other costs next to it. Someone has to read the draft before it goes out, and even a quick, competent check is a minute or two of a real person's time. If the draft is wrong and nobody catches it, the cost is a quote at the wrong price, a booking on a day you can't do, or a customer who got a confident answer to a question they didn't ask. Any one of those is worth more than a year of the model's bill for that task.

The real cost of an agent has three lines

If you are weighing up an AI agent for anything customer-facing, it helps to write the cost down in three parts, because the sales conversation will only ever talk about the first one.

• The run cost. The tokens, the phone minutes, the text messages. Visible, metered, and falling fast. This is what last week's announcements were about.
• The supervision cost. The time someone spends reading, correcting and approving what the agent produced. This doesn't fall when models get cheaper. It falls when the output gets reliably right, and only if you can see that it has.
• The error cost. What it costs when something wrong gets through: the refund, the redo, the awkward phone call, the review. Rare per task, expensive per incident, and the line almost nobody puts a number on.

The honest version of the pitch for any agent is that it moves cost from the second line into the first. It turns a person doing a task into a person checking a task, and the whole case rests on checking being much faster than doing. When that is true, the agent is a genuine win. When the check takes almost as long as doing it yourself — because you can't trust the draft without re-deriving the answer — you have added a step, not removed one.

Why cheaper can make it worse

Cheap runs change behaviour. When each attempt costs a fraction of a cent, the natural move is to point the agent at more things: every enquiry, every follow-up, every review reply, every missed call. The run cost barely notices. The supervision cost scales with the number of things someone has to look at, and the error cost scales with the number of things nobody looked at.

The failure we see most often isn't a bad model. It is a good-enough model deployed wider than anyone can actually watch, so the checking quietly becomes skimming, and skimming becomes approving whatever is on the screen. At that point the business is paying for supervision and getting none.

An agent you can afford to run on everything is not the same as an agent you can afford to be wrong on everything.

What to measure instead

The useful number is not cost per request. It is cost per correct outcome — everything you spent, including people's time, divided by the tasks that ended right. You don't need a spreadsheet model to get a feel for it. Run the agent on one kind of task for a fortnight and keep a tally of three things.

• How many drafts went out unchanged, how many needed an edit, and how many had to be thrown away and done by hand.
• Roughly how long the check took on a typical one. If you find yourself re-reading the enquiry and redoing the maths every time, write that down honestly — that is doing, not checking.
• Anything wrong that got through, and what it took to fix.

If most drafts go out unchanged and the checks are quick, widen the agent's job. If a lot need editing, narrow it — fewer kinds of task, better source information, clearer limits — before you add anything. And if something wrong got through, that is the most valuable line on the page: it tells you exactly where the agent needs a hard rule rather than a better prompt.

Where we land on it

This is the reasoning behind how the AI in Dispatch works: it drafts replies to enquiries from your own price list, and a person still presses send. That is not caution for its own sake. It keeps the supervision cost visible — you can see every draft and how often you change it — which is the only way to know whether the agent is earning its place. We have written before about the difference between software that decides and software that drafts; cheaper models make that distinction more important, not less.

So by all means enjoy the price cuts. They are real, and they make useful things possible that weren't worth the cost a year ago. Just don't let a headline about the smallest line in the sum make the decision about the biggest ones. The question worth asking of any AI agent is not what it costs to run. It is how long it takes you to trust what it did.