Free tool · runs in your browser

Rent the model, rent a GPU or buy one: the break-even

A model billed by the token costs nothing when idle. A GPU costs the same whether it is busy or not. Put in how many tokens you use in a month and today's prices, and see what each of four options costs, the volume at which a machine of your own starts to pay, and how much that answer depends on how full the machine runs.

Nothing you type leaves this page.

Your volume

Tokens a month (millions)Read and written together, from your provider's usage page. A token is about three quarters of a wordParts readInput tokens for each part written. Summarising and answering over documents is about 3 to 1Parts writtenOutput tokens; they cost several times more than input

Paid by the token

Closed model, input ($ per 1M)From the provider's price page, for the tier you would useClosed model, output ($ per 1M)Same page, output columnDiscount you can get (%)Batch work is often half price; cached input far less. 0 for list priceHosted open-weight, input ($ per 1M)A hosting company running a downloadable model for youHosted open-weight, output ($ per 1M)Same page, output columnPrices read onPrices move every few months. Write the date so the result says how old it is

Paid by the machine

Rented GPU ($ an hour)One card of the size your model needsHours billed a month730 is always on. Fewer only if you really switch it offRented GPU, tokens written a secondFlat out, with many requests at once. From a benchmark of your model on that cardYour own hardware ($, once)The card and the machine around itWritten off over (months)How long before you would replace itElectricity ($ a month)Watts times hours times your tariffOwn hardware, tokens written a secondMeasure it with your model. 150 is a placeholder, not a benchmarkUpkeep (hours a month)Updates, monitoring, the evening it stops answering. Counted for both machinesCost of that person's hour ($)Wage plus overheads of whoever looks after it

Where the starting numbers come from

The figures filled in are the ones we read on 12 October 2026 for open-weight models in a small company: a mid-tier closed model at $2 in and $10 out per million tokens, a large open-weight model hosted by the token at $0.15 and $0.60, one 80 GB data-centre GPU at $3.29 an hour, a benchmark of about 2,400 tokens a second on that card under heavy load, a $6,000 workstation, and fifteen hours a month of upkeep. The article links each source. They will be out of date by the time you read this, which is why every one of them is a field: put in today's numbers and the date you read them.

The speed of your own hardware and its electricity have no source. They depend on your card, your model and your tariff, so the values there are placeholders to replace with a measurement.

If you sell software, this bill comes out of gross margin. See what it does next to revenue per head in the margin calculator.

A model, not a forecast: it leaves out the spare machine, traffic that comes in bursts, and the difference in answer quality between models, which you should test on your own work. More free tools: biz59.net/tools.