Calculator · Rod Amora ·
AI Token Cost Calculator
Every token calculator prices tokens. This one also counts the retries, the review, and the correction, because what you are really paying for is a finished piece of work.
You came here to find out what your tokens cost, and that part is arithmetic. Multiply the tokens one task reads by your model's input price per million, do the same for the tokens it writes, multiply by how many tasks you run in a month. The calculator below does it for six current models at their published prices.
Run it and the number will probably come out smaller than you expected. Mine did.
Here is the part the bill leaves out. It counts tokens, so it never sees the retry that fired when the first attempt came back wrong, the four minutes somebody spent reading an output that turned out fine, or the twenty minutes spent fixing one that did not. I cut average token cost on a live workflow about 23 times over and still could not call it a win, because nobody had measured any of that.
So put your own numbers in, and read the second one. Cost per accepted task.
AI token cost calculator
Your token bill
Pick a model, or type the two rates off your own provider bill. Token counts are per task, across every call the task makes.
Claude Sonnet 5: $2.00 per million in, $ 10.00 per million out. Launch rate, now standard.
- Tokens per month
- 7,200,000 in, 1,200,000 out
- Token cost per task
- $0.0528
- Monthly token spend
- $26.40
What the bill leaves out
Retries buy the same token bill twice. Review and correction get paid in somebody's hours. If you have never measured these, put in what you would guess, then watch which one moves the answer.
Cost per accepted task
Cost per accepted task
$7.58
144x the token price
Put your numbers in above. This compares what your provider bills you for one task against what one accepted task really costs you.
- Total monthly cost of this workflow
- $3676.40
- Share that never appears on the bill
- 99%
How AI tokens are calculated
Models do not read words, they read tokens, and a token is a word, part of a word, a punctuation mark, or a patch of an image. English runs somewhere near four characters to a token, so a thousand-word brief is roughly 1,300 of them. If you are budgeting off it rather than guessing, run your own text through your provider's tokenizer, because every model counts a little differently and a rule of thumb is only a rule of thumb.
Billing splits that count in two, and this is the half that surprises people. Input tokens are everything the model reads, which means your prompt, the system instructions, the documents you attached, and the whole conversation so far, again, on every single turn. Output tokens are everything it writes back, and on a reasoning model that includes the thinking you never see.
Output usually costs four or five times input at the same tier. So a workflow that reads a lot and writes a little is cheaper than its token count makes it look, and a chatty agent that keeps re-reading a long thread is more expensive than yours will look on paper.
After that it is two multiplications and an addition. Input tokens over a million times the input rate, plus output tokens over a million times the output rate, times the tasks you run in a month. Caching, batch tiers, and context-length tiers all move that number, which is why the calculator lets you type the two rates you actually pay instead of trusting a list price.
Do AI tokens cost money?
On a paid API, yes, every token, input and output, on every call. On a monthly seat for a chat product you pay a flat fee and the tokens live inside it. Most providers run a free tier with rate limits too, which is enough to test something and not enough to run delivery on.
For a service firm the honest answer is that tokens are the cheap part. Set the calculator to a small model and a normal month and the token spend often lands in the tens of dollars, which is less than one lunch with a client.
I have written the per-task version of this from my own production numbers. A conversation with an agent costs about eight cents at the median, and that cost is a dial you set, not a market price. The bill your firm actually feels is the hours around it.
Cost per accepted task
An accepted task is one that cleared your written standard. First-pass acceptance means it cleared it without anybody fixing it. If the work passes after a correction it still counts as accepted, and the correction time stays in the cost. The tasks nobody ever accepts cost money and produce nothing, so they sit on the top of the fraction and never on the bottom.
Cost per accepted task = (model + tools + infrastructure + retries + review + correction) / accepted tasks
I ran that math on a real decision. I moved a recurring, rule-bound email-to-task workflow onto a cheaper model, and average model cost per attempt went from $1.5474 to $0.0659, about 23 times cheaper, with zero failed attempts on either side. Eleven attempts on the expensive model, ten on the cheap one. The dashboard stayed green the whole way.
I still could not call it a win, and that is the honest part. A green check says the system reached its last step without reporting an error, it says nothing about the work a person still has to do afterward. Nobody had measured whether a reviewer accepted the cheaper output as readily, so what I had was a test result, not a saving. OpenAI's July 14, 2026 guidance says the same thing in its own words, that a cheaper model may need more attempts, more correction, or more human review, and recommends tracking cost per accepted outcome on the workflows that matter.
Once you have the number it belongs next to the others. The Four Numbers covers what changed in the business: cost to acquire, cost to deliver, retention, and price. Cost per accepted task is what feeds the second one. A token bill on its own moves none of them.
Model prices used here
Every rate in the table was read off the provider's own pricing page on August 21, 2026. These are the standard paid rates for text, no batch discount, no cached-input rate, no cache-write premium, no priority tier, because those move the bill in both directions and belong in your own numbers rather than in a preset. Prices move, so check the source before you plan a quarter on one, and if your contract is different, type your own rates in instead.
| Model | Input / 1M | Output / 1M | Source and caveat |
|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | Anthropic · Standard rate. Cache hits bill at 0.1x, batch at half. |
| gpt-5.6-sol | $4.00 | $20.00 | OpenAI · Short-context tier. Long context bills $8.00 and $30.00. |
| Claude Sonnet 5 | $2.00 | $10.00 | Anthropic · Launch rate, now standard. |
| gemini-3.7-flash | $0.75 | $3.75 | Google · Through 2026-12-31, then $1.50 and $7.50. Output includes thinking tokens. |
| gemini-3.5-flash-lite | $0.30 | $2.50 | Google · Output includes thinking tokens. |
| gpt-5.6-luna | $0.20 | $1.20 | OpenAI · Short-context tier. Long context bills $0.40 and $1.80. |
What this calculator does not know
It does arithmetic on the numbers you give it. There is no benchmark inside it and it is not a forecast, so it is only as good as your inputs, and the ones it starts with are round placeholders I picked to look plausible, not figures out of any study.
The bigger limit is that it cannot tell you whether the cheaper model does equivalent work. That is the wall I keep watching firms hit: more capable models make fewer mistakes and cost more, cheaper models push the cost onto human review, and a firm has to measure to find its own sweet spot. There is no table I can hand you for that one.
It also does not model caching, and across my own examples the effective rate swings from about three cents per million tokens to thirteen depending on how hard the system caches. It treats review as a flat average per task when in practice a handful of hard ones eat most of the time. And it says nothing about whether the freed hours turned into anything, which is what the ROI calculator is for.
If you do one thing with the result, take the two numbers in stage 3 to whoever owns the workflow and ask which of those four costs they can actually measure this month. It is usually the review minutes, and the review minutes are usually the biggest slice on the chart. Start there.
FAQ
How are AI tokens calculated?
A token is a chunk of text the model reads or writes, a word, part of a word, or a punctuation mark. English runs somewhere near four characters to a token, so a thousand words is roughly 1,300 of them, though the real count comes from that model tokenizer and shifts with the language. Providers count two kinds. Input tokens are everything the model reads, which means your prompt, the system instructions, the documents you attached, and the whole conversation so far, on every turn. Output tokens are everything it writes back, including the thinking you never see on a reasoning model. The bill is input tokens over a million times the input rate, plus output tokens over a million times the output rate.
Do AI tokens cost money?
On a paid API, yes. Every input and output token is billed on every call, and a long conversation re-reads its own history each turn, so you pay for the same context again and again. A monthly seat for a chat product works differently, you pay a flat fee and the tokens live inside it. Free tiers exist with rate limits, which is enough to test something and not enough to run delivery on. For a service firm the token bill is usually the smallest line in the whole workflow, and the hours around it are the part you feel.
What should an AI token cost calculator actually show me?
Both numbers. Token cost per task is arithmetic and every calculator on the internet gets it right. Cost per accepted task adds the retries, the review time, the correction time, and the platform bill, then divides by the tasks that actually cleared your standard, and that is the one a budget or a model switch should rest on. If a calculator turns a token price into a saving without ever asking who checks the output, it is answering a question you did not have.
Why is cost per accepted task so much higher than the token price?
Because a person costs orders of magnitude more per minute than a model does. Four minutes of review from somebody loaded at sixty dollars an hour is four dollars, and the tokens for that same task often come in under a cent. So the moment a human touches the work, the token line stops being the interesting part of the number. That is not an argument against cheap models. It is an argument for measuring the review and correction time before you switch anything, because that is where the money already is.
Does moving to a cheaper model make delivery cheaper?
Not on its own. I moved one live workflow onto a cheaper model and average token cost per attempt fell about 23 times over, from $1.5474 to $0.0659, with zero failed attempts on either side. I still could not call it a win, because nobody had measured whether a reviewer accepted the cheaper output as readily, and none of the review or correction time was in the comparison. The endpoint here is a tradeoff, not a finish line. More capable models make fewer mistakes and cost more, cheaper models push the cost onto human review, and each firm has to measure to find its own sweet spot.