Prices moved in both directions this month depending on which country the model came from, and this article answers the question underneath all of it: how do you tell whether a cheaper AI model lowers your real cost or quietly raises it?
The Question Everyone Is Asking, And The One They Should Be Asking
The question I get more than any other right now is which AI model is cheapest.
It is the wrong question, and answering it honestly costs people money every week.
I watch business owners switch to a cheaper model, feel smart for about nine days, then quietly spend their Saturday fixing the output nobody caught.
Here is the direct answer. A cheaper model only saves you money if it finishes the job in the same number of attempts with the same amount of your review time. Measure cost per finished piece of work, not cost per million tokens, because a model that costs 60 percent less and fails 40 percent more is not a discount. It is a tax you pay in your own time.
That is the whole thesis. Everything below is the evidence and the system.
Candidly, the reason this question got urgent in August 2026 is that the old shortcut broke.
For two years the shortcut was simple. American models were expensive, Chinese models were cheap, and you picked based on your budget.
That assumption fell apart this month. On August 16, DeepSeek raised API prices on its V4 lineup by anywhere from 50 percent to more than 1,100 percent depending on the model and the hour of day, and introduced peak and off-peak rates. Five days later, OpenAI cut GPT-5.6 Sol pricing by more than 20 percent.
Prices are no longer splitting by capability. They are splitting by geography and by strategy.
Which means the only reliable way to know what you are actually paying is to measure your own finished work.
Key Takeaways
- A cheaper model saves you nothing if it requires more attempts and more of your review time to reach the same finished result.
- Your true number is cost per successful task, which includes model cost multiplied by attempts, plus your own review minutes valued at your real hourly rate.
- The cheap-model-from-China assumption broke in August 2026, with DeepSeek raising prices while OpenAI cut them.
- Defining what "done" means before you prompt reduces retries more reliably than switching to a cheaper model.
- Route work by job type rather than by brand loyalty, and re-check your routing every quarter because prices move faster than habits do.
Why Token Pricing Fools Almost Everyone
Token pricing is the most visible number in AI, and visibility is not the same thing as importance.
You can look up what a model charges per million tokens in about eleven seconds. You cannot look up what it costs you to get a usable client proposal out of it. So people optimize the number they can see.
Here is the thing about that. The number you can see is roughly the smallest line item in the whole operation.
EY's Total Cost of Agents research puts it plainly. A single LLM chat cost about four cents in 2023. An orchestrated agent workflow in 2026 runs about $1.20, roughly thirty times higher, because the workflow now includes tools, reasoning, subagents, planning loops, retries, and refinements before it returns anything to you.
Nothing about that increase shows up in a price-per-token comparison chart.
And that is just the vendor's side of the ledger. Your side is worse, because your side includes you.
Every time a model returns something almost right, you read it, judge it, and decide whether to fix it or run it again. That decision costs you attention, and attention is the scarcest thing a small business owner owns.
I have watched people spend forty-five minutes rescuing a piece of output that cost eleven cents to generate. The eleven cents was never the story.
There is a second trap underneath the first, and it is genuinely counterintuitive. Retrying is not free, and it is not neutral either. When an AI agent fails partway through a multi-step task and you tell it to try again, the failed attempt usually stays sitting in its context window.
Research from Rutgers University published in May 2026, titled "Why Retrying Fails: Context Contamination in LLM Agent Pipelines," measured this against real SWE-bench Verified data. The finding: after a failed attempt, the per-step error rate rises by a factor of 7.1 over the baseline rate.
Read that again slowly. The second attempt is not as good as the first. It is measurably worse, because the failure is still in the room.
The same paper found that standard statistical assumptions overestimate the odds of success across three attempts by 17.4 percentage points, predicting 98.6 percent success where reality delivered 81.2 percent.
So the retry you assumed would rescue the job is less likely to work than you think, and it costs you again. This is why cheap models can be expensive in a way that never shows up on an invoice. You do not get billed for your own Saturday.
I am not telling you to stop caring about price. I am telling you that price without a denominator is not information.
What The August Numbers Actually Show
Look at what the market did this month, and notice that the loudest claims stopped being about benchmarks.
One. The direction of pricing now depends on geography, not capability.
DeepSeek's new rates took effect at 16:00 UTC on August 16, 2026, alongside a peak and off-peak structure where off-peak runs at half the peak rate. V4-Pro output tokens went from a flat $0.87 per million to $3.96 at peak and $1.98 off-peak. The company said the goal was to allocate resources more reasonably.
Meanwhile OpenAI dropped GPT-5.6 Sol input pricing from $5 to $4 per million tokens and output from $30 to $20. That is a promotional rate, held through at least November 21, 2026, not a permanent cut. Worth knowing the difference before you build a budget on it.
One vendor went up. One went down. Same month. Neither move was about the model getting better or worse.
Two. The biggest cost claim of the month was not a price cut at all.
On August 24, OpenAI announced the GPT-5.6 family inside AWS Kiro, and the headline number was an 82 percent cost reduction for completed tasks on Terminal-Bench 2.1. Two caveats worth stating plainly: that figure is specifically for GPT-5.6 Terra, and it comes from joint OpenAI and AWS testing, not an independent evaluator.
Notice the framing. Not a cheaper model. Completed tasks.
And notice where the savings came from. OpenAI credited Kiro's spec-driven approach, where the model receives requirements, design documents, and task context before it starts generating, so it reaches working solutions in fewer iterations and wastes fewer tokens on missteps.
The savings came from better instructions, not a lower sticker price. That is the entire lesson of this article, delivered by OpenAI's own marketing department.
Three. Buyers are already voting with their money, and they are voting for value, not for the top of the leaderboard.
The payments company Ramp analyzed spending across 70,000 businesses and told the Financial Times that Anthropic's flagship Fable 5 accounts for about 11.4 percent of Anthropic dollar spend and only about 6 percent of tokens two months after launch. The cheaper Opus 5, released in late July, overtook it in corporate spending within about a month.
Buyers moved to the cheaper option fast, and they did it inside a single month.
Credit where it is due: I first saw this reported through TLDR AI, and the underlying data belongs to Ramp and the FT.
Four. Serious vendors have stopped selling tokens and started selling outcomes.
Intercom's Fin charges $0.99 per resolution, where a resolution means the customer confirms the issue was solved or does not ask for more help. Zendesk AI agents run roughly $1.50 per automated resolution on a committed plan.
That is the market itself admitting the unit that matters is a finished job.
Five. And here is the sobering backdrop.
MIT's NANDA initiative, in "The GenAI Divide: State of AI in Business 2025," studied 300 AI deployments, interviewed 150 executives, and surveyed more than 350 employees. It found roughly 95 percent of enterprise generative AI pilots produced no measurable financial return.
The report's diagnosis was not that the models were bad. It was a learning gap, meaning companies never integrated the tools into how work actually gets done.
Cheap tokens will not close a learning gap. Better measurement might.
The Finished Work Ledger
Stop comparing models. Start running a ledger.
I want you to think about this the way you would think about a piece of equipment, not a subscription. Nobody buys a truck based on sticker price alone. You ask what it costs to keep it running and how often it leaves you sitting on the shoulder of I-65 with the hazards on.
Same math. Different machine.
Here is the formula, and it fits on an index card.
True cost per finished task = (model cost x number of attempts) + (your review minutes x your real hourly rate) + cleanup cost of anything that slipped through.
That third term is the one people skip, and it is the one that hurts. A wrong number in a client proposal does not cost you tokens. It costs you the client.
Build the system in three layers.
Layer one: the meter.
You cannot manage what you never count. For one week, on one repeating task, write down two numbers every time: how many attempts it took to get something usable, and how many minutes you spent reviewing and fixing it.
Two numbers. A notes app is sufficient. This is not an engineering project.
Layer two: the gate.
Before you prompt, write down what "done" looks like. Length, format, tone, what must be included, what must never appear, and what a failure looks like.
This is the Kiro finding applied to a business with four employees instead of four thousand. The 82 percent came from feeding requirements up front, not from a discount.
Most retries are not model failures. They are specification failures wearing a model's clothes.
Layer three: the router.
Once you have real numbers, assign models to jobs the way you assign people to jobs.
A cheap fast model is genuinely excellent for high-volume, low-stakes, easily verified work: sorting inbound emails, drafting first-pass social captions, summarizing a transcript you will read anyway.
A premium model earns its price on work where a mistake is expensive and hard to spot: client-facing proposals, financial summaries, anything with a legal or medical word in it, anything going out under your name without a second reader.
The mistake is not using cheap models. The mistake is using one model for everything, then arguing about its price.
Tools that help you see this: OpenRouter shows per-request cost across providers in one dashboard, Helicone and Langfuse log spend and tag runs by project, and both Anthropic and OpenAI expose usage breakdowns in their consoles. If you are non-technical and on consumer subscriptions, your meter is a spreadsheet and a timer, and that is completely fine.
[JONATHAN: insert the specific example here of a task you re-measured and what the two numbers turned out to be.]
Seven Steps To Your Real Number
1. Pick one task you do at least weekly. Not your whole business. One repeating job, like weekly client emails, proposal first drafts, or turning meeting notes into action items. You are building a measurement habit, not an accounting department.
2. Write down what "done" means before you touch a model. Format, length, tone, required elements, banned elements, and one sentence describing what a failed output looks like. Keep it in a document you reuse. This single step kills more retries than any model switch.
3. Log attempts and minutes for one week, and price your own hour. Every run of that task gets two numbers: attempts to usable, and review minutes. Then set your hourly value using what you charge clients, or your target annual income divided by 2,000. If you value your hour at zero, every model looks cheap and you will keep choosing badly.
4. Run the same ten jobs through two different models. Same specification, same inputs, same reviewer. Log both. The model that wins on cost per finished task is frequently not the model that wins on price per token, and now you will have proof instead of opinion.
5. Use a structured prompt so the specification travels with the request. Here is the format I teach, the Perfect Prompt Framework 2.0. Keep the four labels separate, never collapsed into a paragraph.
[The Job]
Draft a [TYPE OF DOCUMENT] for [CLIENT OR AUDIENCE] that accomplishes [SPECIFIC OUTCOME].
[The Background]
My business is [WHAT YOU DO]. The reader is [WHO THEY ARE AND WHAT THEY ALREADY KNOW]. Constraints that matter: [LENGTH], [TONE], [ANYTHING THAT MUST BE INCLUDED], [ANYTHING THAT MUST NEVER APPEAR]. Prior versions failed because [WHAT WENT WRONG LAST TIME].
[The Deliverable]
Write it from the perspective of [ROLE OR VIEWPOINT]. Deliver [FORMAT AND STRUCTURE]. Before the draft, list the three assumptions you are making so I can correct them.
[The Questions]
Ask me any questions you have.
6. Start a fresh conversation when a run fails. Do not tell a failed attempt to try harder in the same thread. The Rutgers research shows contaminated context raises the per-step error rate by a factor of 7.1, and clearing context before retrying resolves meaningfully more tasks on the same budget. Copy your specification, open a new window, and go again.
7. Put a quarterly review on your calendar. August 2026 saw one major vendor raise prices and another cut them inside of a week. Ninety days is long enough to be efficient and short enough that you are not paying last quarter's rates out of habit.
Frequently Asked Questions
How do I calculate my AI cost per finished task?
Multiply your model cost by the average number of attempts it takes to get usable output, then add your review and correction time valued at your real hourly rate. Add any cleanup cost from errors that reached a client. That total, divided by finished tasks, is your true number.
Should I switch to a cheaper AI model to save money?
Only after you measure. Run the same ten jobs through both models with identical instructions and log attempts and review minutes for each. If the cheaper model needs more attempts or more of your correction time, the savings are already gone before you finish the week.
Is DeepSeek still worth using after the August 2026 price increase?
It can be, but the calculation changed. DeepSeek raised V4 pricing on August 16, 2026, by 50 percent to more than 1,100 percent depending on model and hour, with off-peak rates at half of peak. If your work can run during off-peak windows, it may still fit. Re-run your numbers.
Do I need technical skills to track AI cost per task?
No. A spreadsheet with three columns, attempts, review minutes, and finished output, gets you most of the value. Tools like OpenRouter, Helicone, and Langfuse add automated tracking if you use APIs, but a non-technical owner on consumer subscriptions can measure this by hand in about ten minutes a week.
Why do AI vendors charge per resolution instead of per token now?
Because outcome pricing aligns what you pay with what you receive. Intercom's Fin charges $0.99 per resolution and Zendesk AI agents run roughly $1.50 per automated resolution, meaning the vendor absorbs the cost of failed attempts. That shift tells you which unit the market considers real.
Buy Finished Work, Not Tokens
Go back to the question people keep asking me: which model is cheapest?
You now have a better one. Which model finishes my work in the fewest attempts with the least of my attention?
That question has an answer specific to your business, and nobody can hand it to you. Not me, not a comparison chart, not a vendor.
But you can produce it in one week with a notes app and a timer, and once you have it, the marketing stops working on you. Price cuts stop feeling like gifts. Price increases stop feeling like betrayals. They become inputs to a calculation you already run.
Here is the most direct thing I will say in this article.
If you are choosing AI tools by sticker price, you are not saving money. You are moving the cost from a line item you can see to one you cannot, which is your own time, and your own time is the one thing in your business you cannot buy more of.
The vendors already figured this out. OpenAI led its Kiro announcement with completed-task cost instead of a benchmark. Intercom and Zendesk sell resolutions instead of tokens. Ramp's data shows corporate buyers already migrating toward value rather than the top of the leaderboard.
They are all measuring finished work. You should too.
If you want the full system, including the specification templates and the routing logic we teach our members, come see what we are building inside White Beard Strategies. The door is open, and the invitation is genuine.
Stop buying tokens. Start buying finished work.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he helps entrepreneurs, coaches, and consultants turn AI from a novelty into a working system. He runs a Facebook community of more than 500,000 members learning AI together, created the Perfect Prompt Framework, and speaks regularly on practical AI adoption for small business. He keeps a running spreadsheet of what his own AI work actually costs per finished piece, and he will tell you honestly that the first month of numbers was humbling.