How much does AI cost? Ask what one finished task costs.

The price of AI fell again this fall, and that tells you almost nothing about what your company will pay. The number to ask for is the cost of one finished task: one invoice read, one customer email answered, one report written. We think the price per token is the wrong number for a business owner to compare, and the monthly subscription price is not much better.

Token prices keep falling

A token is the unit that AI providers bill. A page of about 320 words is about 430 tokens. Providers publish a price for each million tokens that the AI reads and a higher price for each million that it writes.

Those prices keep dropping. On August 21, 2026, OpenAI cut the price of its GPT-5.6 Sol model by 20% for reading and 33% for writing. On September 22 it released GPT-6 Luna at $0.10 for each million tokens read. Its price list shows the model before it, GPT-5.6 Luna, at $0.20. Anthropic had planned to raise the price of Claude Sonnet 5 on September 1. It did not, and the lower launch price is now the standard one.

If you stopped reading there, you would expect every AI bill to go down.

Bills keep rising anyway

On September 4, 2026, the software company Elastic published a piece titled “Why your AI bill tripled while token prices fell 75%”. It says that the blended price per token fell by roughly 75% in a year, and that spending still climbs at companies that run agents. An agent is an AI program that does a job in several steps without a person between them. It looks up an order, checks the policy and then writes the reply. According to the piece, agent work commonly uses 5 to 30 times the tokens of a similar chatbot task.

Elastic sells software for this problem, so treat that as a vendor describing its own customers. But the reason it gives is simple, and we have measured the same thing ourselves. At each step, an agent sends the whole conversation so far back to the AI model, and you pay to have it read again. So a ten-step job means ten readings, and each one is longer than the last.

The plans are changing in the same direction. On September 29, 2026, OpenAI announced a $500 a month plan called Pro 500. On that plan the fastest service tier uses the included allowance first and then draws on a credit balance. GitHub’s coding assistant now counts usage in “AI credits” worth one cent each, and each plan includes a monthly allowance. A flat monthly price increasingly means an allowance, with a meter running after it.

What we measured

We ran this test on a real agent. It used many small tools to do one job, and each step sent the full conversation back to the model. With one of the three models we tested, the median task read 172,078 tokens. That is about 400 pages, read by a machine, to finish one task.

We did not change the AI model. We redesigned the tools the agent uses, so that one call does the work of several and ordinary code handles the bookkeeping. Then we ran the old design and the new design on the same 24 tasks, 595 runs in all, across three models.

The median task read 61 to 73% fewer tokens. With that first model it went from 172,078 tokens to 67,932. Tasks finished in 18 to 22 seconds, where they took 30 to 34 seconds before. The share of tasks that the agent completed went up for all three models.

This was our own benchmark and not an independent audit. We measured tokens and time, not dollars. Even so, compare the size of it. A 20% price cut from a provider is news. A design change on the buyer’s side cut the reading by more than 60%, and nobody would have found it by looking at a price list. One of the price cuts above is promotional, too. OpenAI says the GPT-5.6 Sol price holds “at least through November 21, 2026”. The saving from the design has no end date.

Ask for the cost of the task

So when a vendor quotes you a token price, or says the product runs on the newest and cheapest model, ask a different question. What does one finished task cost, on work like ours? Ask them to count the failed attempts, because you pay for those too.

Then ask what the most is that this can cost in a month, and what happens when it gets there. A product with no cap can surprise you. GitHub, for one, says its spending controls and usage tracking stay available. A control like that only helps if somebody turns it on.

Last, ask who looks at the usage each month. Most companies turn on an AI agent and never check what it costs to run. If nobody at your company is going to look, write that job into the vendor’s contract.

A vendor who can answer all three has measured the product. One who can only repeat the token price has not, and you will be the one who finds out what it costs.

If you already run an AI tool and do not know what a task costs you, book a 30 minute call and we will help you find out.