Because your tools are thinking longer before they answer, and almost nobody is measuring the setting that controls it.
There is a moment in every business owner’s AI journey that arrives without warning. You open the invoice. The number is bigger than last month. And you sit there running the mental tape, trying to remember what you did differently.
You didn’t hire anyone new. You didn’t launch a campaign. You used the same three tools for the same six tasks you always use them for. And the bill went up anyway.
I have had this conversation with more owners in the last ninety days than in the previous year combined. Every one of them assumed the answer was volume. Almost none of them were right.
Here is the direct answer to the question in the headline. Your AI bill is not primarily driven by how much you use it. It is driven by how long your tools think before they answer, and that duration is controlled by a setting most people have never opened. Reasoning tokens, the internal steps a model takes before producing a response, are billed. They are often billed at a premium. And on many tools, the default setting is the expensive one.
That is the whole story. The rest of this article is what to do about it.
Key Takeaways
- Your AI cost is driven more by reasoning effort settings than by how many tasks you run, and most defaults are set high.
- Simon Willison documented a single image generation task consuming 22,276 reasoning tokens over 21 minutes on default settings, versus roughly two minutes with reasoning turned off.
- Agents that retry after a failed tool call burn tokens invisibly, a pattern practitioners now call retry economics.
- DeepSeek raised API prices by as much as 1,100 percent on August 16, 2026, which means promotional pricing is not a stable input cost.
- The metric that protects you is cost per completed task, not cost per month, and you can build it by hand in an afternoon.
You Are Managing a Cost You Cannot See
Most small business owners manage their AI spend the way they manage their electric bill. They look at the total, feel something about it, and move on.
That worked when AI was a flat monthly subscription. It stopped working the moment models started reasoning.
I want to be honest about my own blind spot here. For most of last year I treated my AI spend as a fixed cost. It was a line item. It did not move much. I stopped looking at it in any detail because there did not seem to be any detail worth looking at.
Then I started running agents, and the line item started moving in ways I could not explain from my own behavior. I would have a quiet week and the number would climb. I would have a heavy week and the number would sit still. There was no relationship between what I did and what I paid, which is a fairly reliable sign that you do not understand your own cost structure.
The difficulty here is real and I do not want to minimize it. The information you would need to diagnose this is often not shown to you. Most consumer AI products do not display token consumption. They display a monthly total. You are being asked to manage a variable cost with a fixed cost dashboard.
But what if the reason you cannot find the leak is that you have been looking at the wrong meter entirely?
Four Numbers That Reframe the Problem
One. A single image took 22,276 reasoning tokens and twenty one minutes.
On August 16, 2026, Simon Willison published a hands on review of Qwen 3.8 27B, Alibaba’s newly released open weights model. He praised the model. It runs in 17GB, handles vision, and posts strong coding scores. Then he flagged something that should stop every operator cold.
At the model’s default reasoning setting, generating one SVG image consumed 22,276 reasoning tokens across twenty one minutes. With reasoning disabled, the same task took roughly two minutes. He also noted that a prompt as simple as “draw a circle” triggered an elaborately over engineered animated output.
Nobody chose that. It was the default.
Two. The failure that costs you money is the one that produces nothing.
Practitioner communities on Reddit have started naming a pattern they call retry economics. When an agent calls a tool and the tool returns something unusable, the agent frequently retries. Then retries again. Each attempt consumes tokens. None of them produce output you can use.
You are billed for the attempts, not the results. And because the agent eventually either succeeds or gives up quietly, you often never learn how many attempts it took.
Three. Promotional pricing is not a cost structure.
On August 14, 2026, DeepSeek released V4-Pro. Two days later, on August 16, the company raised API prices by as much as 1,100 percent and introduced peak and off peak tiers, as reported by Caixin Global.
If your margin assumed the old price, your margin was borrowed. Any business built on a single provider’s introductory rate has a cost structure that belongs to someone else.
Four. Most AI projects do not fail on capability. They fail on operations.
MIT research widely reported through 2026 found that roughly 95 percent of enterprise generative AI pilots delivered no measurable financial return. The consistent finding across the coverage was that failures were operational rather than technical. Poor integration with existing workflow, not model malfunction, was the recurring cause.
Cost blindness is an operational failure. It belongs in that 95 percent.
Set those four together and a picture forms that is very different from the one most owners carry. The model is not the problem. The bill is not mysterious. You have simply been running a variable cost process with no unit of measurement, and the variable that moves the most is the one nobody told you existed.
Cost Per Completed Task
The fix is not a tool. It is a unit of measurement.
Stop asking what your AI costs per month. Start asking what it costs per completed job. A job is one finished unit of work your customer would recognize: one client deliverable, one processed order, one drafted proposal, one resolved support ticket.
Here is why that single change does so much work.
Cost per month tells you nothing about whether the number is good. It is one number with no denominator. When it goes up, you cannot tell whether that is growth or waste.
Cost per completed task has a denominator built in. When it rises, one of two things happened. Either you are doing more work, which is fine, or each unit of work got more expensive, which is a problem you can now find and solve.
I built this for my own operation in about an afternoon. It is not sophisticated. I wrote down what one finished unit of work looks like, listed every place an AI tool touches it, and estimated the cost of each touch. The result was a number with a range around it rather than a precise figure, and it was still ten times more useful than the monthly total had ever been.
The first thing it revealed was that my most expensive AI step was one I had never thought about. It was not the writing. It was the research pass that ran before the writing, on a high effort setting, on a task that did not require anything close to that depth.
I turned that setting down. Nobody noticed a difference in the output. That is the entire lesson.
There is a broader principle underneath this, and it is the one I keep coming back to with the entrepreneurs I work with. AI does not remove the need to know your numbers. It creates a new set of numbers to know. The owners who treat AI as a magic box that either works or does not are going to be outrun by the ones who treat it like any other input with a price and a yield.
Practical Steps
1. Define your unit of work before you measure anything.
Write one sentence describing a single completed job in your business. It must be something a customer would recognize as finished. If your unit spans multiple customers or multiple deliverables, it is too big to be useful. Get it down to one thing.
2. Map every AI touch inside that unit.
Walk through the job step by step and mark every point an AI tool is involved. Include the small ones. Especially include the small ones, because small touches at volume are usually where the money is. Most owners find two or three touches they had forgotten entirely.
3. Find and audit the effort settings on every tool you use.
Open each AI tool and look for a control named something like reasoning, thinking, effort, or depth. Note where it currently sits. On most consumer products, the default is set high, because vendors optimize for output quality in a first impression rather than for your monthly bill.
4. Run the three-level test on one real task.
Take one recurring task and run it three times at low, medium, and high effort. Use the same input each time. Then judge the three outputs against a written standard rather than against your gut. Most owners discover that the medium setting is indistinguishable from the high setting for the majority of their work.
Here is the prompt that structures that test properly:
[The Job]
Design a three level effort comparison test for one of my recurring AI tasks.
This is for: me, the owner of [YOUR BUSINESS].
It matters because: I have been paying for the highest effort setting without ever checking whether it improves the result.[The Background]
Here is what you need to know: the task is [DESCRIBE ONE RECURRING TASK IN DETAIL], a good result looks like [DESCRIBE IT SPECIFICALLY], and the customer who receives it is [DESCRIBE THEM].
Do not use: subjective quality language. Give me criteria I can actually check.[The Deliverable]
Return: a test plan under 300 words.
Must include: the exact prompt to run at all three levels, and a five point scoring rubric I fill in myself.
Optimize for: accuracy.[The Questions]
Ask me any questions you have.
5. Set a cost ceiling per task and write down what happens when it breaks.
Pick a number. It does not need to be perfect, it needs to exist. Then decide in advance what you do when a task exceeds it: investigate, downgrade the model, or stop the process. A ceiling with no consequence attached is a wish.
6. Instrument your agents for stop conditions.
If you run anything unattended, give it a time limit and a cost limit. Agents do not know when they are stuck. They will retry until something external tells them to stop, and if nothing does, that something becomes your invoice.
7. Put a fifteen-minute review on the calendar, monthly.
Compare this month’s cost per task against last month’s. That is the entire review. If the number moved and your volume did not, something changed in your process and you now have a month of data to find it in rather than a year.
Frequently Asked Questions
What exactly is a reasoning token?
A reasoning token is part of the internal work a model does before it produces a visible answer. Think of it as the model’s scratch paper. You do not usually see it, but you are billed for it, and on reasoning capable models it is often the largest share of a request’s cost.
Do all AI tools charge for reasoning?
No. Simpler models answer directly with no extended reasoning step, and those are usually much cheaper per task. The tools that carry a reasoning or thinking mode are the ones where the setting matters, and those are increasingly the defaults on premium plans.
Will turning reasoning effort down hurt my output quality?
Sometimes, and only on certain tasks. Work involving judgment, multi step logic, or high consequence decisions genuinely benefits from more reasoning. Formatting, summarizing, and drafting to a template usually do not. The point is to test rather than assume in either direction.
How do I track cost per task if my tool does not show token usage?
Estimate it. Divide your monthly cost by your monthly completed jobs to get a starting figure, then refine it as you learn which steps are heaviest. An estimate with a stated range beats a precise number you do not have.
Is it worth switching to a cheaper provider to save money?
Sometimes, but do the cheaper work first. Most owners can cut their AI cost meaningfully by adjusting effort settings and eliminating retry loops without changing providers at all. Switching adds risk and migration time, so it should be the second move, not the first.
The Close
The invoice that started this article is not really about money. It is about control.
You built a business by knowing your numbers. You know what a customer costs to acquire. You know what an hour of your time is worth. You know your margin, or you know that you should, and you feel the weight of it when you do not.
Then AI arrived and a whole category of cost slipped outside that discipline. Not because you got careless, but because the tools were built to hide it and nobody handed you a meter.
That is fixable this week. Not with a new platform, not with a consultant, not with a course. With one definition, one map, and one number you check on the same day every month.
I would rather you run a smaller AI operation you fully understand than a larger one you are quietly funding on faith. Faith belongs in a lot of places. Your cost structure is not one of them.
Know what a finished job costs you. Everything else is guessing with extra steps.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs turn AI from a novelty into an operating system for their business. He is the creator of the Perfect Prompt Framework, a speaker, and the host of a community of business owners who would rather understand their tools than be impressed by them. He has been running the numbers on his own AI stack since the month the invoice stopped making sense.