7:10 a.m., Atlanta. Denise Carter rolls up the bay door at Carter Auto Spa. Two cars are already waiting. On her phone there's a message from someone who isn't here yet: "Open Sunday?" The answer is one line: "Yes, we're open Sunday morning."
If an AI writes that line, what does it cost? A year ago that was a fair question. Today the answer fits inside the coin on Denise's counter, with change left over.
We took the public price lists of three of the biggest AI model makers and worked out the same thing for each: a thousand short replies to customers.
Carter Auto Spa is our sample business.
📉 Small models got almost eight times cheaper in a year
On October 15, 2025, Anthropic released Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, according to the company's announcement. Today, Anthropic's price list puts its current small model, Claude Haiku 5.5, at $0.10 and $0.50. That's ten times less.
But the same page has a footnote: newer Claude models cut the same text into about 30% more tokens. So the same reply didn't get ten times cheaper. It got almost eight times cheaper. OpenAI's smallest model in its main price table, gpt-6-luna, costs the same $0.10 and $0.50.
A token is a small chunk of text. Google's own guide puts 100 tokens at about 60 to 80 English words. A model cuts into tokens both the question and everything it is given to read alongside it.
On a small model, a simple reply to a customer now costs about an eighth of last year's price. The footnote about how tokens are counted is part of the price too.
📈 Not every price goes down
A year ago, Google charged $0.30 for input and $2.50 for output on Gemini 2.5 Flash, and Google Cloud's price list still shows the same numbers today. The newer Gemini 3.8 Flash on the same page costs $0.75 and $3.75, but that's an introductory price that runs through December 31, 2026. Starting January 1, 2027, the standard price applies: $1.50 and $7.50.
In our math, a reply on Google's newer model costs twice as much as on last year's model today, and four times as much from January.
The fine print has two more lines. The output price covers "reasoning" too: when a model thinks before it answers, you pay for those thoughts at the output rate, even though the customer never sees them. And this model has a free tier, but the same page says that on the free tier, what you send is used to improve Google's products.
Price per word doesn't fall everywhere. An introductory price with an end date is still a price, just one you'll pay later.
🧮 Our math: a thousand replies
Here's the assumption. A reply reads about 2,000 tokens (the question, the start of the conversation, and the business's hours, prices and rules) and writes about 200, which is two or three sentences. That's generous: in our e-mail test with Gemini ("AI on staff," part 1), the real requests were shorter.
- gpt-6-luna today: $0.30 per thousand replies. Claude Haiku 5.5: $0.39. Haiku 4.5 a year ago: $3.00.
- Gemini 2.5 Flash: $1.10. Gemini 3.8 Flash: $2.25 now, $4.50 from January.
- Mid-size models: gpt-6.1-sol $6.00, Claude Sonnet 5.5 $7.80. Sonnet 4.5 a year ago: $9.00.
Each company cuts text into tokens its own way, so comparing companies is approximate. Comparing models of one company is exact.
Today a thousand replies to customers cost between 30 cents and 8 dollars. The difference comes from the model you pick, not from how many customers you have.
What this means for a shop of five
The price of the reply itself has almost left the bill. An owner of a car wash or a salon has no reason to count tokens. Two other prices are still there.
The first is what you pay to make sure there's a reply at all: a subscription, a service, or your own evening. Divide it by replies, not months, and only count the replies that actually went out and were right.
The second is a reply that never went out, or one that was wrong. No price list counts that one. If nobody answers "Open Sunday?" at 7:10 a.m., the car goes on Sunday to whoever did.
What to take with you
- Count last month's customer questions as one number: messages, calls and e-mails together.
- Divide what you pay for answers (a service, a subscription, your own hours) by the replies that went out and were right. That's your real price per reply.
- Ask your provider what counts as one reply and what happens when the monthly limit runs out.
- Find the end date of any introductory price you rely on. For Gemini 3.8 Flash, it's December 31, 2026.
- Once a week, read ten replies in a row and mark the wrong ones. One wrong reply costs more than a thousand tokens.
How many customer questions reach you in a week, and how many of them wait until morning for an answer?
