API or your own GPU: the break-even, explained
How we compute the tokens-per-hour above which renting a card beats paying per token for an open-weight model, what goes into the number, and what it deliberately leaves out.
The question
For a model with open weights you have two ways to run it. Pay an API per million tokens and pay nothing while idle, or rent a GPU by the hour and pay the same whether it is busy or not. Below some volume the API is cheaper; above it the card is. The page called API or your own GPU computes that crossover for every open-weight model we track, using the prices we collect the same day.
The arithmetic
Three inputs, all measured. First, the API price of the model on OpenRouter, in dollars per million tokens, blended from the input and output prices assuming three tokens in for every token out, which is a typical chat ratio. Second, the smallest GPU the model fits on at 16-bit precision with a 32,000-token context, from the same memory arithmetic the GPU pages use: parameters times two bytes, plus the attention cache, plus a 15 percent margin. Third, the cheapest hourly price for that card on the providers we track that day.
Break-even is the hourly price of the card divided by the blended API price. The result is in millions of tokens per hour: the throughput at which the rented card costs exactly what the API would charge for the same tokens. Push more tokens through than that per hour and renting wins; fewer, and the API wins.
What the number does not say
It does not say whether your setup reaches that throughput. Tokens per hour on a given card depend on quantization, batch size, context length and the serving stack, and none of that is something we have measured. A break-even of 10 million tokens per hour is arithmetic; whether your server produces 10 million tokens per hour is a benchmark, and the page says so in the caveat above the table.
It also assumes one card. When a model does not fit on any single GPU we track, the page shows how many of the cheapest fitting card it would take, and the hourly cost multiplies accordingly. Multi-card setups have their own overheads that the arithmetic ignores.
Why prices change the answer
Both sides move. API prices on OpenRouter change when providers reprice or when a model gains or loses a free tier; GPU prices move with marketplace supply. Because we recompute from the day's snapshots, the same model can cross from API-wins to renting-wins in a week without anything changing in your workload. The daily reports call out those moves when they happen.