Anthropic shipped a model on September 28 that runs about 30 percent faster at the exact same price, which raises the question this article answers: how do you decide which job goes to which model, and what is it costing you to keep sending everything to the same one?
A client told me last week that her AI bill had tripled and she could not explain a single dollar of it.
She was not being reckless. She picked a tool in early 2025, liked it, and never looked back. Everything went through the same model. Subject lines, meeting notes, a vendor contract, a forty page strategy document. One door for every job.
Then Anthropic released Claude Sonnet 5.5 on September 28, and that one door got a lot more expensive to keep using.
So here is the direct answer to the question in the title. Stop asking what AI costs per month and start asking what it costs per completed task. Route your high volume, low judgment work to a fast mid tier model, keep your most expensive model for the work where a mistake compounds, and you will usually spend less and finish faster at the same time.
That practice has a name. Model routing. It is the skill almost nobody is teaching, and it separates an AI bill that scales with your revenue from one that scales with your inattention.
What makes this urgent is the specific thing Anthropic shipped. Per Anthropic's own announcement page for Claude Sonnet 5.5, the model is roughly 30 percent faster than Sonnet 5 and carries identical pricing: two dollars per million input tokens, ten dollars per million output tokens, twenty cents per million cache reads. TechCrunch's September 28 coverage makes the mechanism plain. The savings Anthropic claims, up to 30 percent, do not come from a price cut. They come from the model needing fewer tokens and fewer tool calls to finish the same work.
Read that again, because it is the whole argument. The price on the page did not move. The cost of getting something done did.
If your budget line says "AI tools, $200 a month," you just got handed a real improvement and you will never see it, because you are measuring the wrong number.
Key Takeaways
- The number that decides your AI economics is cost per completed task, not cost per month and not cost per token.
- Anthropic priced Claude Sonnet 5.5 identically to Sonnet 5 at two dollars per million input tokens and ten dollars per million output tokens, so the savings come entirely from finishing work in fewer tokens.
- Kilo's six month study of Reddit developer conversation found that switching, comparing, and combining models outranked every other theme, including pricing and reliability.
- Ramp's spend data shows a twenty times cost spread between the cheapest and most expensive models businesses actually run, which makes model selection a finance decision, not a technical one.
- McKinsey's 2026 survey found 80 percent of respondents report AI improved their own productivity while only 37 percent report any EBIT impact, and that gap is a routing problem as much as a strategy problem.
The Problem Is That You Picked Once and Never Picked Again
Here is the thing about how most business owners ended up where they are.
You did not make a bad decision. You made one decision, at one moment, with the information available then, and the world kept moving while your workflow stood still.
Somewhere in 2024 or 2025 you tried a few AI tools, one of them clicked, and you built habits around it. Those habits turned into templates. The templates turned into how your team works. Nobody revisits a decision that is not visibly broken.
Meanwhile the field shipped a new frontier model every few weeks, the naming conventions turned into a cruel joke, and the benchmarks published alongside each release measured tasks that have nothing to do with drafting your follow up emails.
Candidly, the industry made this hard on purpose. When every lab claims the best model, the only honest response from a busy owner is to pick one and stop paying attention. I have made that call myself.
But look at what it produces. You end up running your highest volume, lowest judgment work on your most expensive option. Every subject line, every meeting recap, every "clean up this paragraph" request goes to the model you chose for the hardest problem you ever faced.
That is like buying a truck rated to tow twelve thousand pounds so you can haul a bag of mulch home from the garden center twice a week. The truck is not wrong. Using it that way is.
And it is not only about money. A slower, heavier model on routine work costs you the thing you actually cannot replace, which is the twenty minutes you sat there watching it think.
The reframe is simple and it takes about an hour to install. Stop treating your AI as a subscription you own and start treating it as a set of workers you assign. Some jobs need your best person. Most jobs need your fastest person.
The Evidence That Routing Beats Loyalty
I do not want to argue this from opinion, so here is what the record actually shows.
Developers already solved this, and they solved it by routing. Kilo published a study on September 3, 2026, authored by Job Rietbergen, analyzing 6,876 Reddit posts and comments across 266 subreddits from January through June of this year. They went in looking for which model developers liked best. What they found instead was routing: developers assigning different models to different parts of the job, mostly because the alternatives kept costing them money or time. Switching, comparing, and combining models beat every other theme in the corpus, including pricing and reliability.
The failed attempt is the hidden line item. In that same Kilo study, reliability was the most negative theme measured. Four hundred thirty four items, 31.8 percent of that theme, carried explicitly negative language, making it the only major theme where negative sentiment crossed 30 percent. Kilo's own summary of the math developers kept running is the sentence I want on your wall: a cheap request that needs three attempts costs more than an expensive one that finishes on the first try. Note that this cuts both directions. Routing down is not the same as buying cheap.
The price spread between models is enormous, and it is a finance decision. Ramp published 2026 token spend benchmarks drawn from its Token Spend Management product, which sees what businesses actually paid Anthropic, OpenAI, and other providers. Their observed effective rates for April 2026 ranged from seven cents per million tokens at the low end to $1.42 per million at the high end, a twenty times spread. Ramp's framing is exactly right: model tier is now a finance decision, and engineering books the tickets while finance gets the bill. In a small business, you are both people.
Most businesses run far more models than they realize. Ramp's same data shows the median business used nine models in April 2026 while the average used 16.5. Among the highest AI adopters on Ramp, median monthly token spend was $2,246 against an average of $140,842, and that gap tells you one unattended workflow can move a company from typical to alarming. Worth saying plainly: that median comes from Ramp's heaviest adopters, not from a typical small business, so treat it as the shape of the problem rather than your number.
The productivity is real while the profit is not showing up. McKinsey's State of AI survey, published August 25, 2026 and fielded to 1,719 respondents across 97 nations between May 4 and June 8, found that 80 percent of respondents say AI improved their individual productivity. Only 37 percent report AI contributing to their organization's EBIT, essentially unchanged from 2025. McKinsey also reports that about one in five organizations say AI operating costs, including tokens, have constrained their AI use.
Eighty percent feel faster. Thirty seven percent can find it in the numbers. That gap is partly strategy, and it is partly this: nobody is measuring the unit that would let them see the improvement.
What Changed for Me: The Cost Per Completed Task Test
Here is the reframe that fixed this in my own business.
I stopped asking "what does this tool cost" and started asking "what does this task cost, start to finish, including the times it did not work."
The formula is not complicated. Cost per completed task equals the cost of one attempt, times the number of attempts it takes to get something usable, plus your time cleaning up what came back.
That third term is the one everybody skips, and it is usually the largest. If a model saves you eight cents and costs you eleven minutes of editing, you lost.
Once you write it that way, the routing decision makes itself. You are sorting your work by how much judgment it requires and matching each pile to the cheapest option that reliably finishes on the first try.
I sort into three piles.
Reflex work. High volume, low judgment, easy to verify at a glance. Subject lines, meeting summaries, reformatting, first draft social captions, turning notes into an outline. You can tell instantly if it is wrong. Send this to your fastest mid tier model and never think about it again. This is exactly the lane Anthropic positioned Sonnet 5.5 for: well scoped everyday work, bug fixes, polished documents, slides, and spreadsheets.
Judgment work. The output is hard to verify quickly and a mistake propagates. Pricing analysis, a contract you are about to sign, a strategic plan your next six months rest on, anything where being confidently wrong costs you real money. This is where you pay for the heavier model, which Anthropic positions as Opus 5.5. Paying more here is not indulgence. It is insurance.
Comparison work. The decisions worth getting a second opinion on. Kilo's data showed developers routinely splitting planning from implementation across different models rather than picking one for everything. You can do the same thing without any technical setup. Cuey, which launched on Product Hunt this week with 255 upvotes, puts ChatGPT, Claude, and Gemini answers side by side in one browser tab. That is a two minute habit that catches the confident wrong answer before it becomes your decision.
The point is not that one lab wins. The point is having a system instead of a default.
How to Build Your Routing Map This Week
-
List your last twenty AI tasks before you change anything. Open your chat history and write down what you actually asked for, not what you think you use AI for. Most owners are surprised by how much of it is reflex work. You cannot route a workload you have never looked at.
-
Sort each task into reflex, judgment, or comparison. One question decides it: if this comes back wrong, do I notice immediately or three weeks from now? Immediate means reflex. Three weeks means judgment. Anything where you would want a second opinion from a smart friend goes in comparison.
-
Move your entire reflex pile down one tier and run it for a week. Do not migrate everything at once. Take the biggest pile, the boring one, and send it to the faster mid tier model. Track two things only: did it finish on the first try, and how long did it take you to fix.
-
Measure attempts to done, not satisfaction. Satisfaction is a feeling and it lies. Attempts to done is a number. If the cheaper model takes two tries where the expensive one took one, you did not save anything and you should route that task back up. Kilo found that exact pattern across thousands of developer conversations.
-
Turn on caching or reuse your context deliberately. Ramp's data shows businesses paid an effective 62 cents per million tokens on Claude Sonnet 4.6 in April 2026, against its three dollar list price, and the gap was caching. Practically, that means stop re pasting the same background into every new chat. Use a project, a custom instruction, or a saved brief so the reused context gets cheap.
-
Put a named owner on every recurring AI workflow. If nobody owns it, nobody notices when it quietly starts costing three times what it did in July. Ramp found that month to month AI spend swings a median of 58 percent in absolute terms across the businesses it tracks, and that 61 percent of businesses exceed 40 percent month over month volatility. Traditional software subscriptions vary by almost nothing. This does not, and most owners find out after the invoice.
-
Re run this audit quarterly, not at every launch. Cognition cut Devin's pricing 30 to 40 percent in the same window Anthropic shipped Sonnet 5.5, and Google restructured its consumer AI tiers. Chasing each announcement is a job. Checking your routing four times a year is an hour.
Here is a prompt to run step one and two in a single pass. Fill the brackets with your real details.
[The Job]
Help me build a model routing map for my business so I stop sending every task to the same AI model.
[The Background]
I run a [TYPE OF BUSINESS] with [NUMBER] people. The AI tasks I do most often are: [LIST 8 TO 12 RECURRING TASKS]. Right now I send all of it to [MODEL OR TOOL] and pay [AMOUNT] per month. The models I have access to are [LIST THEM].
[The Deliverable]
A table with one row per task and four columns: the task, how much judgment it needs rated low, medium, or high, which model I should send it to, and one sentence explaining why. Below the table, name the three tasks where switching saves me the most, and tell me the specific warning sign that would mean I routed one of them wrong. Written from the perspective of an operations lead speaking to a non technical business owner.
[The Questions]
Ask me any questions you have.
Frequently Asked Questions
How do I figure out my cost per completed task if I pay a flat monthly subscription?
Use time instead of dollars. Count how many attempts a task takes and how many minutes you spend fixing the output. A flat subscription hides the money but not the minutes, and on a subscription plan your real currency is usage limits and your own attention, both of which a wrong model burns fast.
Is the cheapest model always the right choice for simple work?
No, and this is where people lose money. Kilo's study found that a cheap request needing three attempts costs more than an expensive one that finishes first. Route down only until the first try stops succeeding reliably, then stop. The goal is the cheapest model that finishes cleanly, not the cheapest model.
Do I need technical skills to route models?
Not anymore. Most major AI apps now let you pick a model from a dropdown before you type. Routing is a decision you make in two seconds, not a system you build. If you use several tools, a side by side comparison app like Cuey handles the second opinion without any setup at all.
Should I switch providers every time a new model launches?
No. Switching has a real cost in rebuilt prompts, lost context, and retrained habits. Review your routing quarterly instead of chasing announcements. The exception is when your current provider changes pricing or limits in a way that affects your highest volume workflow, which is worth checking within a week.
Why did my AI bill go up when the prices went down?
Because volume grew faster than price fell, and because premium model usage crept upward. Ramp found premium models moved from 5.7 percent of business AI cost in June 2025 to 55.9 percent in April 2026. Your team quietly upgraded tiers for quality reasons and nobody flagged it as a spending decision.
The Close
My client's bill tripled because she never had to make a second decision. One model, every job, for eighteen months. That is not a spending problem. It is a system that was never designed, just inherited.
You are probably in the same position, and it is not a character flaw. It is what happens when a field moves faster than anyone's capacity to keep re deciding.
But the ground shifted on September 28. Anthropic put a better model behind the same price tag, and the only way to capture that is to know which of your jobs belong there.
So do the unglamorous thing. Open your chat history. Write down twenty tasks. Sort them into reflex, judgment, and comparison. Move the reflex pile down a tier and watch what happens to your week.
Eighty percent of the people McKinsey surveyed feel faster. Thirty seven percent can prove it in the numbers. The difference between those two groups is not a better model. It is knowing which model each job deserves.
Stop buying AI by the month. Start buying it by the task.
About the author
Jonathan Mast is the founder of White Beard Strategies, where he teaches non technical entrepreneurs to use AI in ways that survive contact with a real business. He runs a Facebook community of more than 500,000 members and the AI Insiders membership, and he speaks regularly on practical AI adoption for small businesses. He runs his own operation on a routing map exactly like the one in this article, and he rebuilt it three times before it held.
If you want the routing audit run alongside people doing it at the same time, that is the kind of work we do inside AI Insiders, our membership at $247 per month. Our Express AI Training sessions run $10 live and $20 for the replay if you would rather start smaller.
