OpenAI’s chief financial officer has handed buyers the right metric and kept the pen for herself. Sarah Friar’s scorecard, published on July 17, tells you to judge AI by cost per successful task, and defines it to include staff time, review, retries and rework. Good. Then her worked example prices only the API bill, the one line her own definition says not to stop at.
Steal the idea. Throw out the example. Build your own before the next renewal.
Key points
- What is cost per successful task? It’s the total spent to get a piece of work finished to your standard, divided by the number of pieces that met the standard.
- Why can the cheaper model cost more? Because the saving on the API bill is small next to the cost of a person fixing errors.
- Who gets to write the scorecard? Whoever picks the benchmark picks the winner, and that holds for every vendor.
- Why are flat plans turning into meters? Look at what happened to Anthropic’s Fable 5 the same week.
What is cost per successful task?
It’s the total spent to get a piece of work finished to your standard, divided by the number of pieces that met the standard. The spend includes the model, retries, the person who reviews the output and the rework when it’s wrong.
Friar’s scorecard lists four measures, which are useful work, cost per successful task, dependability, and value as usage grows.
That beats comparing rates per million tokens. A token price tells you what the model charges for trying. It says nothing about how often the try is good enough, which is why our look at GPT-5.5’s doubled price per token told readers to price the finished task instead. Friar’s page makes the same point, that the lowest price per token doesn’t always give the lowest cost per outcome once retries and review are added.
Why can the cheaper model cost more?
Because the saving on the API bill is small next to the cost of a person fixing errors. A model that fails slightly more often sends slightly more work to a human, and human minutes cost far more than machine tokens. The sum flips the moment review time enters it.
Think of two delivery vans. One runs on cheaper fuel and has a habit of arriving with the wrong parcels. The fuel saving is real and the reshipping eats it. Here’s the same logic as a hypothetical, ours and not a measurement of any model. Say 1,000 tasks a month, and each failure takes a person 20 minutes, which at $50 an hour is about $16.67.
| API cost per task | Right first time | Fix cost per task | Total per finished task | |
|---|---|---|---|---|
| Model A | $0.50 | 90% | $1.67 | $2.17 |
| Model B (36% cheaper) | $0.32 | 86% | $2.33 | $2.65 |
The cheaper bill produces a finished task that costs about 22% more. Model B wins the line item and loses the month.
Change the assumptions and the answer moves, which is the point of building your own. The break-even sits at a fix cost of $4.50 a failure, or roughly five and a half minutes at $50 an hour. Any fix that takes longer than that and Model B costs more per finished task. Most real fixes, like re-reading a contract summary or rewriting a customer reply, run well past five minutes. Our piece on the hidden time tax of AI looks at the same problem from the staff side.
Who gets to write the scorecard?
Whoever picks the benchmark picks the winner, and that holds for every vendor. Friar illustrates the method on DeepSWE v1.1, a coding test, where OpenAI’s Sol scored 72.7% against 69.9% for Claude Fable 5, at an estimated API cost 36.2% lower.
Those numbers come from OpenAI, on a test of its choosing, and they cover the API bill. For scale, OpenAI’s launch rates were $5 in and $30 out per million tokens for Sol, and $1 and $6 for the cheaper Luna tier. The human-review half of her own formula isn’t in the example.
None of that makes the example false. It makes it a vendor’s illustration of a method, and a sensible buyer treats it that way. Your tasks aren’t software engineering tasks. A scorecard built from your own invoices and support tickets will beat any lab’s example, including ours.
Why are flat plans turning into meters?
Look at what happened to Anthropic’s Fable 5 the same week. Anthropic said Claude Fable 5 would stay on its Max and Team Premium plans at up to 50% of their weekly usage limits, with Pro and Team Standard subscribers getting a one-time $100 credit before moving to usage credits priced roughly at API rates, according to PCWorld.
The Decoder reports that Anthropic had first planned to pull Fable from subscription plans entirely, and PCWorld says the company extended that deadline more than once before reversing. Our read is that plan caps are turning into metered billing, and a flat subscription is slowly becoming a variable cost. A variable cost needs a unit price you can watch, and cost per successful task is the unit.
Don’t expect a token price cut to shrink the total either, as we argued in our June piece on OpenAI price cuts and the Uber budget.
How do you build your own scorecard?
Take one workflow and write down five numbers. Tasks per month. Plan or API cost per task, with overage. The share of outputs that need a human fix. Minutes per fix, priced at the loaded hourly rate of whoever does it.
Then add the plan or API cost per task to the fix share times the cost of a fix, and you have your cost per successful task.
Update it each quarter. A two-week sample of 30 tasks with a stopwatch is enough. Staff will guess at fix times, and the sceptic will say that makes the whole sum fuzzy. Fair, but a guess still moves the ranking in the example above, because the human cost dwarfs the token cost. Per-token comparison feels precise and answers the wrong question. Our guide to AI costs for Canadian business shows what the rest of the spend looks like.
Then use it where money changes hands. A vendor quoting you a cheaper rate has to explain how many extra fixes the cheaper rate brings. A vendor pushing a price rise has to explain what it does to the finished-task number, not the token number. Once you hold that figure, the conversation moves onto ground you chose. Without it, you’re comparing rate cards the seller formatted.
The hypothetical numbers are ours, and OpenAI’s 36.2% figure is its own estimate on one coding benchmark, so it says little about writing, analysis or customer support. Prices and caps also changed several times in the month before this article. OpenAI later cut Luna’s price by 80% and Terra’s by 20% on July 30, ten days after this article’s date, while Sol’s price held. A moving price list is one more reason to anchor on the finished task.
Resist the temptation to build a model that’s too tidy. The first version should fit on a single sheet with five inputs and one result, and it should be wrong in ways you can see. A rough number that staff understand will be updated. A precise one that only finance understands will be filed and forgotten, and then the vendor’s example is the only scorecard in the room.
The one exception is output that goes straight to customers with no review at all. Then the API bill really is the cost. Everyone else should write the five numbers for their most-used workflow this week and bring them to the next renewal.
Frequently asked questions
What is cost per successful task?
It’s total cost, including staff review and rework, divided by the number of tasks that met your quality bar. It replaces price per token as the number to compare.
Is the cheaper AI model always cheaper?
No. If it needs more human fixes, the saving on the API bill can disappear. In our hypothetical, a 36% cheaper bill still cost about 22% more per finished task.
How do I calculate AI cost per task?
Add the API or plan cost per task to the share needing a fix times the cost of a fix. Use a loaded hourly rate and a sample of at least 30 real tasks.
Archive entry dated July 20, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
