AI API Cost Calculator: The Same Chatbot Costs $30 or $1,370 a Month

Every model provider publishes a price per million tokens, and every one of those numbers is useless on its own. Nobody thinks in millions of tokens. You think in “about a thousand people a day ask my bot something”. So here is the thing that translates between the two.

Move the numbers below to match your actual usage. The prices are the real published rates, checked on 4 September 2026, and you can edit any of them if something changes.

The spread is the whole story

Run the default numbers, a thousand requests a day with a small prompt and a short answer, and you get a range that should stop you in your tracks. The cheapest option lands near 30 dollars a month. The most expensive lands near 1,370. That is the same work, done by ten models that all speak fluent English and all pass the benchmarks their makers care about.

A 45x price gap between the cheapest and the priciest model is not a rounding error in your budget. It is the difference between a side project and a line item that needs approval.

What surprised me most was not the top of the range. It was how tightly the middle clusters. GPT-5.6 Terra and Gemini 3.1 Pro are priced identically, two dollars in and twelve dollars out. Claude Sonnet 5 sits just under them. When three providers converge on the same number to the cent, that is not coincidence, that is a market rate forming.

The prices, verified

Two notes on this table, because both caught me out. Gemini 3.5 Pro is still not on Google’s price list, so the delay I wrote about in July has not resolved. And the GPT-5.6 numbers you find on comparison blogs are frequently wrong, because OpenAI cut prices on 30 July 2026 and a lot of pages never updated. Luna dropped by 80 percent in that cut. If a site is quoting Luna at a dollar, it is quoting you last quarter.

Output tokens are what actually bill you

Look at the ratio in that table. Output costs five to six times what input costs, on almost every model, from every vendor. That means the single biggest lever on your bill is not which model you pick. It is how much text you let it produce.

Set the calculator to a thousand input tokens and a hundred output tokens, then flip it to a hundred in and a thousand out. Same total tokens, wildly different bill. I have watched people agonise over a model choice that moves their cost by 20 percent while shipping a prompt that says “explain your reasoning in detail” and triples it.

The caching discount nobody puts in the spreadsheet

Drag the cache slider up and watch the GPT-5.6 rows fall off a cliff. Cached input on those tiers bills at 10 percent of the normal rate. If you are sending the same long system prompt on every request, and most production apps are, that is most of your input cost gone.

I want to be straight about a limitation here. I only apply the cached rate to models where I could verify a published cached price, which right now is the three GPT-5.6 tiers. Other providers offer caching too. So a high cache setting in this tool flatters OpenAI, and you should read it as “what caching can do” rather than a fair head to head.

What I actually do, and why it is not about the money

I should be honest about something. Privately I do not pay per token at all. I pay a flat monthly subscription, so nothing in this calculator changes my bill by a cent.

It still changed how I work. I used to run everything through the top model because it felt like the safe choice. In busy weeks I kept hitting my usage limit, and hitting a limit in the middle of a task is its own kind of expensive. Now I match the model to the job: the strongest one for research and anything where facts matter, a mid tier for long running conversations where it is plainly good enough, and a small fast one for rewording and polish. Same habit the calculator argues for, different currency. Flat rate users are not buying dollars, they are buying headroom.

At work it runs the other way round. There I am billed per API call, so the same habit turns straight back into money. Which is why I get slightly evangelical about this. If you are working on a company key, your employer is paying for every “what is the weather” and “what should I have for lunch” you push through the expensive model. Do not be that person. Send the cheap questions to the cheap model.

Same price does not mean same experience

Here is something the table cannot tell you. GPT-5.6 Terra and Gemini 3.1 Pro cost exactly the same, two dollars in and twelve dollars out. That does not make them the same product.

I want to be careful, because my own Gemini time has been on the consumer subscription rather than the API, so this is not a controlled comparison. For the way I work it has been the weaker fit. Answers came back simpler than I wanted, I found myself asking follow up questions to get to the depth I needed, and in longer sessions it lost the thread more often than I liked. Steering it felt harder. Where it has genuinely won me over is images and video, where my Adobe credits let me generate without a watermark, and that part I use happily.

The point is not that one vendor is bad. It is that when two models cost the same to the cent, price stops being the deciding input and your own testing has to start. Which is the whole argument for running your real workload through a model for a week before you commit to it.

How I would actually use this

Pick the cheapest model in the list, not the best one. Run your real workload against it for a week. If the quality holds, you are done and you just saved 90 percent. If it breaks, you now know exactly what it breaks on, which is the only information that justifies paying more.

The mistake is starting at the top. People reach for the flagship because it feels safer, then never test down, and they carry that decision for a year. The calculator is there to show you what that habit costs.