Start with the price list, because the price list explains the habit. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. At the Bank of Canada rate of 1.4123, that is about $5.65 CAD and $28.25 CAD. Every major provider bills the same way: by the token.
A token is a chunk of text, roughly a short word. Providers count tokens, price tokens, and report tokens back to you. So your dashboard fills with token counts. The held detail is who now says that number misleads. The chief executive of the largest seller is one of them.
The direct answer: tokens measure computation, not completed work. A rising token count can mean a workflow is producing more. It can also mean the model is wandering, retrying, and restating. The invoice cannot tell those two apart. Measure finished outputs instead.
- The meter is real, the unit is not
- What the seller said about his own unit
- Amazon wrote the warning down
- Why dashboards love tokens anyway
- The one-week test
The meter is real, the unit is not
Nobody disputes the meter itself. Providers really do process tokens, and the processing really does cost money. The dispute is about what the count means for a buyer. A token total says how much text moved through the machine. It says nothing about what your firm got.
Two teams can burn identical token counts in a month. One shipped forty client reports and a new proposal system. The other spent the month re-prompting a chatbot that drafts badly. Their dashboards look the same. Their quarters do not.
Canadian teams already fight a version of this problem. Claude Personal Finance Canada is still US-only. Ban bank CSV pastes first covers tools whose headline metrics hide local limits. Open Google Vids before you buy Gemini Omni API makes the same point on price. The visible number is rarely the deciding number.
What the seller said about his own unit
On October 31, 2026, Sam Altman discussed pricing on a podcast with Stripe co-founder John Collison. He called tokens a kind of fake unit. His fuller point was precise: tokens are real as a measure of use. They are fake as a measure of business value.
He also explained why the unit survives anyway. Customers prefer a bill they can predict, even a slightly inflated one. Offered a choice, buyers pick the countable thing over the honest thing. They would rather overpay inside a unit they understand. That preference, not accuracy, keeps token pricing in place.
Sit with that for a moment. The seller offers the metric because buyers demand it. Buyers then report the metric to their boards as progress. The loop flatters everyone and informs no one.
Amazon wrote the warning down
Amazon reached the same conclusion from inside a big engineering org. Its updated engineering guidelines, published October 27, 2026, warn against tokenmaxxing. That is the practice of inflating AI usage to look productive on internal dashboards.
The guidelines call usage metrics deeply flawed when disconnected from results. Adoption rates, token consumption, and code change counts all get named. The company had already shut an internal usage leaderboard in May. Employees were gaming it to climb the rankings.
Amazon’s replacement yardstick is long-term outcomes. Not how much AI the team consumed. What the team delivered, and whether it held up. That is a buyer’s metric, not a vendor’s metric.
Why dashboards love tokens anyway
Tokens are easy to count, easy to chart, and easy to bonus. Finished work is none of those things. Counting outputs means defining a finished unit for each workflow. A signed report. A resolved ticket. A proposal that reached a client.
There is a second, quieter cost hiding under the token count. Teams optimise what the dashboard shows, so they generate more text. The extra text still needs human review before it can ship. Review AI-generated work with one signed line before your name goes on it describes that review burden plainly. More tokens can mean more unpaid checking, not more output.
The prompt layer hides a similar leak. You already paid for your AI tools. In hours, rewriting the same prompts prices the hours teams burn re-deriving instructions. Token dashboards never show that cost either. Your real AI bill has three lines, and the invoice shows one.
The one-week test
No hands-on test ran for this piece, so here is the test, named and dated. Pick one workflow your team runs every week. Count its finished outputs for five working days, starting Monday. A finished output is one a client or a manager accepted without a rebuild.
Write the token total beside the output count each day. On Friday, divide one by the other. That ratio is your first honest AI metric. If outputs rose while tokens fell, your process improved. If tokens doubled and outputs held still, your process is leaking.
Then put one question to your vendor in writing. Ask for cost per finished task on your top workflow, not cost per million tokens. The answer, or the silence, tells you which side of the metric your vendor is on. Report the ratio to your board instead of the token chart.
