Is a local LLM cheaper than an API?

Per token, easily. Renting a million tokens of Qwen3.8 27B costs 45.6c. Generating the same million on a Mac Studio M5 Max, 64GB costs 1.7c of electricity, 26 times less. The machine costs $3,499, which is 7.97B tokens of that gap, and that number is the whole argument.

Run the numbers on that pairing

Where those two figures come from

A million tokens at 15:1 is 937,500 you send and 62,500 the model writes back. The API bills both. At $0.32 per million in and $2.50 per million out, that is 30.0c for the input and 15.6c for the output: 45.6c the million. Most of the bill is the context you send, not the answer you get.

At home you pay for the electricity the machine draws while it writes those 62,500 tokens. At 24.7 tok/s that is 2,525 seconds, about 42 minutes of generation; at 145 W it draws 0.102 kWh; at $0.17 per kWh that is 1.7c.

The input is counted in the million but costs no generation time here, because prompt processing runs far faster than generation. It is not free, so read the electricity figure as a floor rather than a final number. The 145 W is a stand-in: nobody has put a meter on this machine, and its own page says what the figure borrows. Double it and a million tokens still costs 3.5c to generate against 45.6c to rent. Electricity is at United States prices (US EIA).

A million tokens, model by model

Every model a Mac Studio M5 Max, 64GB holds at 32k context, priced both ways. The last column is what matters: how many tokens have to go through the machine before the gap has covered the $3,499.

26 of the 27 models here pay the machine back at some point. LFM2.5 2.6B does not, because it is listed free by the cheapest host on the date checked, and a free endpoint cannot be beaten on price. A model nobody rents is priced as its closest hosted match and the row says so.

The token count holds still. The date moves.

Break-even on this site is usually a number of months, and months depend on how hard you work the machine. Counted in tokens it does not: the saving on each million is the same at any level of use, so the count that covers $3,499 is the same too. Here is Qwen3.8 27B on a Mac Studio M5 Max, 64GB at the five levels the calculator names.

That is the case for buying, and the case against it, in one table. The machine is cheap per token and expensive to own, so the only thing that makes it pay is volume you actually have.

What you use it for moves the line further than what you buy

The mix matters because the two sides bill it differently. The API charges for every token you send; the machine spends its time and its watts on the tokens it writes. So work that sends a lot and writes a little is the cheapest to rent, and the slowest to justify a machine.

Writing and drafting pays the machine back in 2.2B tokens; search over your own documents needs 8.52B. Buying the hardware changes less than that. Across the 30 current machines that run Qwen3.8 27B at 32k, a million tokens costs between 1.6c and 4.3c to generate. Their prices run from $1,269 to $18,000.

Each row is a mix the calculator offers, and the one marked default is what every other figure on this site is priced at: whole files and tool results resent on every turn, which is why it sends the most.

Where renting still wins

Of the 949 machine-and-model pairings this site can price, electricity beats the API on 892. On 37 the model is listed free by the cheapest host, so there is nothing to beat. The remaining 20 are all one model, Devstral Small 2 24B. Nobody rents it, so it is priced as gpt-oss-20b, which moves 3.6B parameters for every token against this one's 24B. It is dense, so it is slow on every machine that holds it, and the price it borrows is one a far lighter model sets.

Renting also wins any time you would not have used the machine. A pay-back counted in billions of tokens is not a prediction that you will reach them: at 200k tokens a day, light assistant use, the count arrives in 109 years, which is a long way of saying never. The reasons to buy that survive that table are the ones this arithmetic does not price: your data staying on your desk, no rate limit, no outage, and a model that still runs when the endpoint is retired.

Every figure is Qwen3.8 27B at Q4_K_M unless the row names another model, at 32k context, 15:1 input to output, $0.17 per kWh, and today's API prices held flat. Rental prices have not held flat: they have fallen steeply and repeatedly for a given level of capability (Epoch AI). The calculator can decay them 40% a year, and switching that on pushes every count on this page out. The speed above is worked out from memory bandwidth rather than measured, and each machine page says which of the two it has. To change any of this, open the calculator, or see what pays back soonest at each level of use.