Self-hosted LLM vs API: break-even calculator

100/121 prices match the official page, 96 match models.devSynced 2026-11-19 01:00 UTC · Method

A rented GPU server costs the same every hour whether it is busy or idle; an API bills only the tokens you send. Enter your server and workload to see which is cheaper and from how many requests per day self-hosting pays off.

API models vs your server

API models vs your server
ModelProvider APISelf-hostedBreak-even req/day

Server cost = GPU price per hour × GPUs × hours per day × 30. One server carries throughput × seconds on × utilization tokens per month (input + output). Break-even = requests per day at which the API bill equals one server; — means the API is cheaper per token even on a fully used server.

How the estimate works

The throughput you enter decides the result: measure it with your model, GPU and serving stack (for example vLLM) at your real prompt lengths. Leave room for peaks with the utilization figure, and add engineering time, which the server price does not include.

FAQ

When is self-hosting an LLM cheaper than an API?

When the server stays busy. A GPU bills every hour, so the cost per token falls as utilization rises; at low volume a cheap API model almost always wins, and many small API models stay cheaper per token than any rented GPU.

What throughput should I enter?

Total tokens per second (input + output) the whole server sustains with many requests in parallel, measured with your model and prompt sizes. Input tokens are processed much faster than output, so long prompts with short answers give a higher figure.