Every major AI provider now sells roughly the same job at three or four prices. The gap between the cheapest and the most expensive isn’t 20 percent. On the models listed today it’s 200 times on input and 125 times on output.

A small business usually buys the top one by default. Not because anyone chose it, but because it’s the option the software picked before you ever saw a price.

What the tiers actually cost

Prices per million tokens, taken from the providers’ own pricing pages in October 2026. A million tokens is roughly 750,000 words, so read these as “per large pile of text” rather than per request.

Model Input Output
GPT-6-Astra (OpenAI, top tier) $10.00 $50.00
Claude Opus 5.5 (Anthropic, top tier) $4.00 $20.00
Claude Sonnet 5.5 (Anthropic, mid) $2.00 $10.00
Gemini flagship reasoning (Google) $2.00 $10.00
GPT-6-Luna (OpenAI, budget) $0.10 $0.50
GPT-5-Nano (OpenAI, cheapest) $0.05 $0.40

The top and bottom rows are both OpenAI models, sold by the same company, on the same account, in the same month.

One job, priced two ways

Take a real job: incoming email gets read and sorted. Say you handle 2,000 enquiries a month, and each one means the model reads about 2,000 words and writes a two-line summary. That’s roughly 4 million input tokens and 400,000 output tokens a month.

On GPT-6-Astra: 4M in at $10, plus 0.4M out at $50. $60.00.

On GPT-5-Nano: 4M in at $0.05, plus 0.4M out at $0.40. $0.36.

Same job, same month, same company. The difference is $715 a year, and most of it buys nothing.

Where the money actually goes

Output tokens cost 5 to 8 times input tokens on every model in that table. That’s the number most people miss.

Sort your work by what it produces, not what it reads:

  • Reading and sorting. Classifying enquiries, tagging invoices, routing a form, deciding which of three inboxes an email belongs in. Short answers, so cheap on both ends.
  • Reading and writing a little. Summaries, first drafts of a paragraph, rewriting a product description. Output starts to matter.
  • Reading and writing a lot. Reports, long emails, anything where the useful part is the text coming back out. This is where bills get away from you.

A job that reads 2,000 words and answers “yes” or “no” is almost free on any model. A job that reads 2,000 words and writes 2,000 words on a top-tier model is where you start noticing the invoice.

The five-minute test that settles it

Ignore benchmarks. They measure someone else’s tasks.

Take the ten requests you send most. Run each one through three models: a top-tier one, a mid one, and a budget one. Read the nine answers with no idea which model made which. Mark anything wrong, vague, or confidently incorrect.

Three results turn up almost every time:

  1. Most tasks score the same on all three. Sorting, summarizing, extracting dates, rewriting for tone. This is usually 60 to 80 percent of a small business’s AI use. Downgrade these and nothing changes except the bill.
  2. A few need the expensive model. Anything with judgment, multi-step reasoning, or nuance about your specific trade. Keep these where they are.
  3. A few fail on every model. These aren’t AI tasks. They need a written rule, a form field, or a person.

That third group is the one nobody expects and the one that saves the most time, because it stops you paying to automate something that should be a checklist.

The trap in the other direction

Switching everything to the cheapest model is not the answer. A model that gets 3 percent of your classifications wrong is not cheaper. It is a model plus a review step, and someone has to do the reviewing.

Price the mistake, not the token. If an error means a customer waits three days or an invoice goes out wrong, the mid-tier model is the cheap one.

What to do this week

Find out which model your tools are calling. In ChatGPT, Claude, or Gemini’s own settings, look for model selection. In any tool that claims to “have AI built in”, ask support directly: which model, and can I change it. If they cannot answer, that is your answer.

Then run the ten-task test. It costs a few dollars and an afternoon, and it decides your running cost for the next year.

If you would rather someone else ran it and put the numbers beside a payback figure, that is what a discovery call is for.