In June 2026, TechCrunch rounded up the chaos companies faced with AI bills: Uber used its entire 2026 AI coding budget by April; one company’s routine contract renewal came back at four to five times the price; a CTO said a single engineer spent $40,000 on tokens in one month. J.R. Storment, executive director of the FinOps Foundation, which promotes cloud financial management, said the conversation had shifted from “go fast” to “we need guardrails, how do we control this?”
The report also cited an observation: the heaviest users were about twice as productive as light users but used ten times the tokens. Whether that spend pays off depends on the business value of the code that ships, which most companies still can’t measure.
That’s the blind spot of comparing model prices alone. The price per million tokens is the vendor’s rate card; your cost is something else. The real cost of a task includes how many retries it took, how much material was fed in, and how much time people spent fixing the result before it was accepted. To manage an AI bill, start by calculating what each accepted result cost.
A cheaper model doesn’t guarantee a smaller bill
Rate cards really do differ. Anthropic’s official pricing page, for example, prices output tokens several times higher than input, while reused cached content is much cheaper. Amazon Bedrock and other cloud platforms likewise price by model and usage pattern. So the same task can cost very different amounts depending on how it’s handed to AI.
But switching to a cheaper model can raise total cost if quality drops, retries increase, and people spend more time fixing results. Conversely, a pricier model that gets it right the first time may be cheaper overall. Unit price alone hides all of this.
Calculate a per-task cost for common work
Pick three to five tasks your team most often gives AI, such as meeting summaries, code changes, or draft customer replies, and log each for a week or two:
Tokens spent: input volume, output length, and number of retries. Retries and refeeding the whole document each time are often the main reasons a bill grows.
How many were accepted: how many results were actually used, and how many were discarded and redone.
Time people spent: how long people still had to fix and check results before accepting them.
Divide total spend by the number of accepted results to get cost per result; add the human fix-up time and compare with doing the work without AI. That number tells you which tasks deserve a stronger model, which should first narrow the input and pin down the output format, and which don’t suit AI at all.
Adjust the task design before arguing about budget
Once you have per-task costs, common adjustments include: chunking or summarizing material before handing it to AI instead of feeding the whole thing; specifying output fields and length instead of “as complete as possible”; and capping retries, so that if two or three attempts in a row fail, you stop and check whether the problem was defined clearly.
Long-running tasks burn money especially fast, mostly through retries and rereading context; see Long AI Tasks Spend Most of Their Budget on Retries and Rereading Context. For team-level budget caps and handling overages, see With Copilot on Usage-Based Billing, Decide First Whether to Stop or Pay When Credits Run Out.
The value a per-task cost can’t capture
Per-task cost only captures the visible part. Some value is hard to quantify, such as getting new hires up to speed faster or letting one person handle work they couldn’t before. These can go into the evaluation but shouldn’t be forced into a number.
Also, logging has its own cost. You don’t need to measure every task; the few most frequent and most expensive ones are enough. The industry is still developing ways to measure this, too; the report notes the Linux Foundation is organizing an effort to define metrics for AI cost.
AI handoff card
Our team's AI bill keeps growing, and I want to calculate what each accepted result actually costs for a few common tasks, not just compare model prices.
The method comes from this article: https://boosterminiclass.com/en/posts/model-cost-guardrails-before-ai-token-bill/
Its view: the price per million tokens is a rate card; real cost includes the number of retries, how much material is fed in, and the time people spend fixing results before they're accepted. Calculate the cost per accepted result first, then decide whether to upgrade, scale back, or stop.
First ask me: the three to five tasks the team most often gives AI, the models or tools used, and what usage data I can currently see.
Then:
1. Design a two-week log: token usage per task, number of retries, whether it was accepted, and human fix-up time;
2. Explain how to calculate "cost per accepted result" from the log, including people's time;
3. List common adjustments, narrowing input, limiting output, capping retries, and which kinds of tasks each suits;
4. Give decision rules: when a stronger model is worth it, when to redesign the task first, and when a task isn't suited to AI.
Mark anything without real numbers as "needs logging," and don't calculate with assumed figures.
References
- TechCrunch: The token bill comes due: Inside the industry scramble to manage AI’s runaway costs — https://techcrunch.com/2026/06/05/the-token-bill-comes-due-inside-the-industry-scramble-to-manage-ais-runaway-costs/ (2026-06-05)
- Anthropic official documentation: Pricing — https://platform.claude.com/docs/en/about-claude/pricing [accessed: 2026-10-11]
- AWS official: Amazon Bedrock Pricing — https://aws.amazon.com/bedrock/pricing/ [accessed: 2026-10-11]
