AiGanakSelf-hosted model or API: which is cheaper for you?
Compare what you would pay an API provider with what running your own model on rented GPUs costs, and see the volume where self-hosting starts to pay off.
A free calculator that compares the monthly cost of an LLM API (OpenAI, Anthropic or Google, at list prices checked weekly) with running your own open-weight or custom small language model on rented GPUs, and shows the break-even volume. Inputs: requests, tokens, GPU price, throughput, utilisation and operations cost.
Monthly cost as volume grows
How the comparison works
API cost = requests × (uncached input × input price + cached input × cached price + output × output price), at the provider's list prices per million tokens (batch prices if you tick batch).
Self-hosted cost = GPUs × hourly price × 730 hours + your monthly engineering and operations cost. GPUs needed = the larger of your minimum and the GPU-seconds your traffic needs (input tokens ÷ input speed + output tokens ÷ output speed, per request) divided by the seconds in a month at your utilisation, rounded up.
Break-even is the monthly request volume where the API bill passes the self-hosted bill. Self-hosting is a step function: each extra GPU adds a fixed cost, so the lines can cross more than once.
What it leaves out: answer quality (a self-hosted model must be good enough for your task), the one-off cost of fine-tuning or building a model, storage and network, and volume discounts on either side. Treat the result as a first estimate.
Self-hosting vs API questions
Can't find it? Press ⌘K and ask.
When is self-hosting an LLM cheaper than an API?
When your volume is high enough that the API bill passes the fixed cost of GPUs plus the people to run them. With small, efficient open-weight models and steady traffic this often happens at millions of requests a month; with large models or spiky traffic it may never happen. The calculator finds the break-even point for your numbers.
Where do the GPU price and throughput defaults come from?
They are example assumptions, not quotes: $2.00 per GPU hour and 1,500 output and 15,000 input tokens per second per GPU, roughly a 7-8B model with batched serving on one data-centre GPU. Replace them with your cloud provider's price and throughput you measure on your own hardware.
What does the calculator leave out?
Answer quality, the one-off cost of fine-tuning or training a model, storage, network, and volume discounts. A self-hosted model only saves money if it is good enough for your task, so test it on your own examples first.
Does AiGanak build self-hosted models?
Yes. AiGanak builds custom small language models and deploys open-weight models in a company's own cloud or on-premise, starting with a proof of concept on the company's data. Book a call from the contact page.
Thinking about your own model?
We build and deploy custom small language models in your own cloud or on-premise, after a proof of concept on your data. Book a 30-minute call to check whether it pays off for you.