Is the cheaper model always cheaper?

Only below the break-even retry rate. A model at a third of the price that needs two attempts on half of requests has given the saving back.

When does self hosting pay?

Above a volume threshold, and only if you exclude nothing. The infrastructure is the visible cost and the engineering time to run it reliably is usually larger.

Technologyai

AI API Cost Comparison Calculator

Model cost comparison with retries counted. A cheaper model that needs a second attempt on a fifth of requests is not a fifth cheaper, and the break-even retry rate is what decides the choice.

Also called: llm pricing comparison, model cost comparison.

Cheaper option

Model B is cheaper

Model B is cheaper. Model A costs ₹1,75,000 a month and model B ₹56,000 including a 0% retry rate, a difference of ₹1,19,000 or 68%. Model B stops being cheaper above a 212.5% retry rate.

Model A

₹1,75,000

Model B

₹56,000

Model B before retries

₹56,000

Difference

₹1,19,000

Difference as a percentage

68%

Self hosting

₹2,50,000

Retry rate at which they cost the same

212.5

On retries

Model B stops being cheaper above a 212.5% retry rate.

On prices

Prices are inputs, not constants. Model pricing changes often, so take the current rate from the provider rather than relying on a default here.

How this is calculated

Method and background

How this is calculated

Both models are priced on the same token volumes. The retry rate is the input that makes this honest: a cheaper model that fails or produces unusable output some of the time costs its rate card plus the retries, and often plus a fallback to the expensive model. The break-even retry rate shown is where the two become equal, which is usually a lower number than people expect. Self hosting is a fixed cost against a variable one, so it wins only above a volume threshold and carries operational cost that no rate card shows.

a cheaper model that needs retries is not as cheap as its rate card suggests

r: Retry rate

Worked examples

Each of these is asserted on every build. If a change to the engine ever moved one of these answers, the build would fail before the page could print it.

a cheaper model with no retries

Requests a month: 2,00,000
Average input tokens: 1,500
Average output tokens: 400
Model A input per million: ₹250
Model A output per million: ₹1,250
Model B input per million: ₹80
Model B output per million: ₹400
Extra retries needed on the cheaper model: 0%
Compare against self hosting: No
Self hosting monthly cost: ₹2,50,000

Cheaper optionModel B is cheaper

300M in at 250 plus 80M out at 1250

Open this example

retries erode the saving

Requests a month: 2,00,000
Average input tokens: 1,500
Average output tokens: 400
Model A input per million: ₹250
Model A output per million: ₹1,250
Model B input per million: ₹80
Model B output per million: ₹400
Extra retries needed on the cheaper model: 50%
Compare against self hosting: No
Self hosting monthly cost: ₹2,50,000

Cheaper optionModel B is cheaper

boundary: half the requests attempted twice

Open this example

Method and limits

What it assumes

Identical token volumes on both models, which different models rarely produce.

What it deliberately does not model

Different models produce different output lengths for the same task, which this holds constant.
Quality differences are not captured by any cost comparison.
Self hosting cost excludes engineering time, which usually dominates.

Formula version 1.0.0 · definition 1.0.0 · India · Report a problem with this calculator

Frequently asked questions

Is the cheaper model always cheaper?: Only below the break-even retry rate. A model at a third of the price that needs two attempts on half of requests has given the saving back.
When does self hosting pay?: Above a volume threshold, and only if you exclude nothing. The infrastructure is the visible cost and the engineering time to run it reliably is usually larger.