The question this article answers: if per-token price is not what determines your AI costs, what actually does, and how do you fix it?
Hook and Direct Answer
I watched a client stare at an invoice with the exact expression of a man who has just been told his flight is cancelled. He had done everything right. He had read the pricing pages. He had compared input and output rates across four providers. He had migrated his whole workflow to the cheapest model he could find. And his monthly bill had gone up 40 percent.
He looked at me and said, “Jonathan, I don’t understand. I picked the cheap one.”
Here is the direct answer. You are not billed on price. You are billed on price multiplied by tokens. And the model you picked to save money is burning three times as many tokens to finish the same job, because it is less capable, so it flails, backtracks, re-reads, second-guesses, and takes twenty-six turns to do what a better model does in eleven. The cheapest model on the pricing page is almost never the cheapest model on your invoice.
I have been where you are. I spent months treating per-token price as the lever, negotiating and comparing and switching, and getting nowhere. Here is what changed for me: I stopped asking “what does this model cost per million tokens” and started asking “how many round trips does it take to get to done.”
Per-token price is an input. It is not an outcome. What you actually pay is decided by how many tokens your architecture burns to finish a job. That means the fix is not procurement. The fix is architecture. Stop negotiating price. Start reducing the number of round trips it takes to get to done.
That is the whole thesis, and the rest of this piece is the proof and the playbook.
Key Takeaways
- Your AI bill is price multiplied by tokens consumed, and tokens consumed is a function of your architecture, not the provider’s rate card.
- A model that costs twice as much per token can produce a smaller invoice if it finishes the job in fewer turns and drags less context along.
- Cognition proved this in July 2026: a model at 2x the per-token price came in cheaper per completed task than the model it replaced, and scored higher.
- The metric that matters is cost per completed job, not cost per million tokens. Most entrepreneurs have never measured it once.
- Gartner predicts over 40 percent of agentic AI projects will be cancelled by the end of 2027, with escalating costs named as a leading cause.
The Problem
Here is the honest version of what is happening, and I want to say it plainly because most people selling you AI advice will not.
The pricing page is the only number that is easy to see. So it is the number everybody optimizes. It sits right there in a clean table, two columns, input and output, dollars per million tokens. You can compare it in thirty seconds. It feels like the decision.
It is not the decision. It is one term in a multiplication problem, and you have no visibility into the other term.
The other term is how many tokens your workflow actually consumes to finish one real job. Not one prompt. One job. The whole thing, start to shipped. And that number is invisible to you unless you go looking for it, because nobody puts it on a marketing page. It depends on how many times the model goes back and forth. How much conversation history it drags into every single call. Whether it re-reads a file it already read. Whether it does work itself that it could have handed to something cheaper. Whether it second-guesses good work and redoes it at premium rates.
I got this wrong for a long time. I want to be honest about that, because I think the pretending is what keeps people stuck. I built a research workflow last year that I was genuinely proud of. Smart prompts. Good outputs. And it cost me a small fortune per run, and I blamed the model. I switched to a cheaper one and the bill barely moved, because the cheaper model needed more attempts. So I switched back and blamed the provider. Then I actually sat down and read the logs, call by call, and found that my workflow was re-sending the same 18,000 token brief on every single turn, and taking about thirty turns per run. The model was never the problem. My architecture was the problem. I was paying to say the same thing thirty times.
This is difficult work. I will not pretend otherwise. Reading call logs is not glamorous, and there is no dashboard that hands you the answer. But here is the reframe, and it is a good one. If your bill is set by architecture rather than by price, then your bill is something you control. Price is a thing that happens to you. Architecture is a thing you decide. That is not bad news. That is the best news in this entire industry right now, because it means the lever is in your hands and almost nobody is pulling it.
The Evidence
I do not want you to take my word for any of this. Here is the research.
Cognition ran the experiment and published the receipts. On July 13, 2026, Cognition, the company behind the Devin coding agent, published “Making Fable Cheaper Than Opus.” They replaced Claude Opus 4.8 with Fable 5 inside Devin. Fable 5 costs twice as much per token. Their bill went down. They ran 3,000 evaluation sessions on the FrontierCode 1.1 benchmark across four configurations. The result: Fable plus a cheap sidekick model cost $1.86 per run and scored 60.7. Opus plus the same sidekick cost $2.04 per run and scored 54.6. Cheaper and better, from the model with double the sticker price. Their own words: “The 2x per-token premium turns out to be the wrong number to look at.”
The mechanism is turns and context, not price. In the same study, Cognition parsed every LLM call in all 3,000 sessions. The Fable lead took 11.5 turns per run. The Opus lead took 26.5. Fable’s lead consumed 545,000 cumulative input tokens per run. Opus consumed 1,679,000. That is roughly three times the input, from the model that costs half as much per token. Both models delegated the same number of times, about three handoffs per run. The difference was when and what they delegated. In 81 percent of Fable-led runs, the lead model never made a single code edit itself. For Opus, that was true in only 24 percent of runs.
Architecture beats model choice, and the same study proves it twice. Compare pure Fable to Fable plus a sidekick. Pure Fable scored 60.8 at $4.03 per run. Fable plus a sidekick scored 60.7 at $1.86. Same score, 54 percent less money. The score did not move. The architecture did. That is the cleanest illustration I have seen of the point I am making: you did not change the model, you changed the shape of the work, and you cut the bill in half.
Token consumption, not model choice, predicts outcomes. Anthropic published an engineering breakdown of their multi-agent research system showing that token usage alone explained roughly 80 percent of the performance variance they observed, and that multi-agent systems burned about 15 times more tokens than a normal chat. Read that from the cost side. If tokens drive both your results and your bill, then tokens, not price, are the thing to design around.
The market is already voting, and most people have not noticed. Vercel’s AI Gateway Production Index for July 2026 reported that open-weight models handled 29 percent of gateway token volume in June while accounting for less than 4 percent of spending. Almost a third of the work, a twenty-fifth of the money. That gap is not a pricing story. That is a routing story. Somebody decided which jobs needed a frontier model and which did not.
And the stakes are real. Gartner predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Escalating costs is first on that list. These projects are not dying because the models are bad. They are dying because nobody measured what a finished job actually costs until the invoice arrived.
The Solution
Here is what changed for me, and it is embarrassingly simple.
I replaced one metric with another. I stopped tracking cost per million tokens. I started tracking cost per done.
Cost per done is exactly what it sounds like. Pick one real job your business actually needs finished. A client onboarding packet. A researched blog draft. A month of social content. A cleaned-up CRM segment. Then measure what it costs, end to end, to get that one job to shipped. Not per prompt. Per job. That single number is the only AI cost metric that has ever changed a decision I made.
Once you have it, the framework I teach inside White Beard Strategies has four moves. I call it the Cost Per Done Framework, and it spells DONE, which is the point.
Define done. Most AI cost overruns are actually specification failures wearing a costume. If the model does not know what finished looks like, it will keep going, and every extra turn is money. Cognition found this exact thing. When Fable delegated, it wrote briefs that listed constraints, edge cases, and an explicit definition of done. On one benchmark task, that discipline was the difference between a score of 25 and a score of 94, on identical work. Vague requests are not just lower quality. They are more expensive. Write the definition of done before you write the prompt.
Observe the trajectory. Open your logs. I know. Do it anyway. Count the turns it takes to finish one real job. Count how many times the same context gets re-sent. Look for the moment the model re-reads something it already had. You cannot fix a round trip you have never counted, and I promise you the count is higher than your guess. Mine was thirty when I would have sworn it was eight.
Narrow the lead’s job. This is the big one. Your most expensive model should be doing judgment, not typing. Deciding what to build, setting constraints, reviewing the result. Everything mechanical goes to something cheap. This is the entire Cognition finding in one sentence: the frontier model earns its price when it is managing, and burns your money when it is doing. In 81 percent of their winning runs, the expensive model never touched the code. Ask yourself what your expensive model is currently doing that a cheap one could do with a good brief.
Evaluate against cost per done. Now you can actually make decisions. Change one thing. Rerun the job. Compare cost per done. Not vibes, not the pricing page, the number. This is where you discover things that feel like they should not be true, like a pricier model producing a smaller bill.
I run this on my own stack every quarter. It is not exotic. It is the same discipline any operator applies to any input cost, applied to a line item that most entrepreneurs have decided is somebody else’s department. It is not somebody else’s department. It is yours, and it is one of the few places left where a small business can outperform a company with fifty times the budget, because they are still comparing rate cards.
Practical Steps
Pick one job and price it end to end. Choose a real, repeatable job with a clear finish line. Run it once and capture total spend from first call to shipped. That number is your baseline, and it will probably surprise you.
Count your turns. Go into your logs and count how many round trips that job took. Write the number down where you can see it. Every turn you eliminate is money you keep, forever, on every future run.Kill the context you are re-sending. Look for the same brief, the same instructions, the same history riding along on every call. Turn on prompt caching if your provider offers it. Anthropic’s prompt caching cuts costs on cached content by up to 90 percent and latency by up to 85 percent on long prompts. This is an afternoon of work for a permanent discount.
Write a real definition of done. For your one job, write down the constraints, the edge cases, and what finished looks like, before the model starts. This is the highest leverage unpaid work in your whole stack. It buys quality and cheapness at the same time, which almost nothing else does.
Demote the mechanical work. Identify every step in the job that is typing, formatting, extracting, or summarizing rather than deciding. Route those to a cheap model. Keep the expensive model for judgment and review only. Start with one step, not all of them.
Rerun and compare cost per done. Same job, new architecture, new number. If it went down, keep it. If it went up, you learned where your judgment actually lives. Both outcomes are wins. Only the unmeasured version is a loss.
Make it a quarterly ritual. Model prices, capabilities, and tiers are changing every few weeks right now. GPT-5.6 shipped in three separate price tiers on July 9. Your optimal architecture in October will not be your optimal architecture today. Put it on the calendar.
Frequently Asked Questions
Is the cheapest model ever actually the cheapest choice?
Yes, for narrow mechanical jobs with a clear finish line, like extraction, formatting, or classification. Cheap models win when the task offers no room to flail. They lose on open-ended work where capability reduces turn count. Match the model to the job’s shape, not to your budget anxiety.
How do I measure cost per done without engineering help?
Run one real job start to finish and read the total from your provider’s usage dashboard. Divide by jobs completed. That is it. Most platforms show spend by day and by key. If you can create a separate API key per workflow, you can get this number this afternoon without writing a line of code.
Does this apply to me if I just use ChatGPT or Claude on a subscription?
Yes, in time rather than dollars. Round trips are still round trips. If it takes you fourteen exchanges to get a usable draft, you are paying in your own hours, which cost more than tokens. Better specifications and clearer definitions of done shorten that loop the same way.
Are open-weight models worth looking at for a small business?
Increasingly, yes. Vercel reported that open-weight models handled 29 percent of AI Gateway token volume in June 2026 while accounting for under 4 percent of spending. GLM-5.2 leads open-weight models on the Artificial Analysis Intelligence Index at roughly one sixth the price of GPT-5.5. Use them for volume work, not judgment work.
What is the single biggest waste in a typical AI workflow?
Re-sending the same context on every turn, followed closely by letting an expensive model do mechanical work it should have delegated. Cognition’s data showed one lead model consuming three times the input tokens of another on identical tasks. That is not a pricing problem. That is the same words, purchased repeatedly.
The Close
I want to go back to my client and his invoice, because I think about that face a lot.
He did not make a dumb mistake. He made the only mistake available to him. He optimized the number he could see. That is what good operators do. The whole industry hands you a pricing page and lets you believe that is the decision, and then quietly bills you for an architecture nobody ever showed you how to look at.
So let me hand you the thing I wish someone had handed me two years ago. Your AI bill is not a price you accept. It is a design you chose, mostly by accident. Every round trip in your workflow is a line item. Every re-sent brief is a purchase order you signed without reading. Every time your best model does work your cheapest model could have done, you paid frontier rates for typing.
And this is why I get genuinely fired up about it. Because 40 percent of agentic projects are going to get cancelled by 2027, and most of those teams will blame the models. They will not be right. They will have been paying for round trips they never counted, and calling it the price of AI. Meanwhile a company with double the sticker price and half the turns walked away cheaper and better, and published the receipts, and almost nobody read them.
You are not too small for this. You are exactly the right size for this. The advantage does not go to whoever negotiated the best rate. It goes to whoever needs the fewest attempts to get to done. That is a discipline, not a budget, and disciplines are free.
If you want the full framework, the log audit walkthrough, and the templates we use to write definitions of done that cut turn counts in half, that is what we build together inside White Beard Strategies. Come get it. Bring your ugliest workflow.
Stop shopping for a cheaper model. Start building a shorter road to done.
About the Author
Jonathan Mast is the founder of White Beard Strategies, where he serves a community of more than 100,000 entrepreneurs learning to use AI as a business advantage rather than an expensive toy. He is the creator of the Perfect Prompt Framework, a widely used method for getting reliable, repeatable results out of AI systems, and a sought-after speaker on practical AI implementation for small and mid-sized businesses. Jonathan’s work focuses on one stubborn idea: that the entrepreneurs who win with AI are not the ones with the biggest budgets, but the ones who build the clearest systems. He writes and teaches from experience, having built, broken, and rebuilt his own AI stack more times than he cares to admit.