Why compare prompt and completion separately?

Providers price output tokens several times higher than input, so the two levers have different leverage. The analyzer varies each independently so you can see whether shaving the prompt or capping the output gives the bigger reduction for your workload.

When does output length dominate the bill?

When completions are long relative to the prompt, or whenever the model has expensive output pricing. In those cases setting a tight max_tokens and asking for concise answers usually beats prompt trimming.

When does prompt length dominate?

For retrieval-augmented or long-context tasks where you stuff large documents in and ask for a short answer. There the input tokens drive cost, so compression, caching and chunking pay off most.

Is my data sent anywhere?

No. All calculations run in your browser. Nothing you enter is uploaded, stored or logged.

Cost Sensitivity Analyzer

Cost sensitivity analyzer

Should you spend effort trimming the prompt or capping the completion? It depends on which one your bill is sensitive to. This analyzer takes your baseline tokens and model, then sweeps prompt and completion length independently across a range so you can see exactly which lever moves total cost more — and stop optimizing the one that barely matters.

How it works

The cost of one request is:

cost = (input_tokens / 1,000,000) × input_price
     + (output_tokens / 1,000,000) × output_price

The tool holds one value at its baseline and varies the other from −range to +range (e.g. −50% to +50%), then repeats with the roles swapped. Because output is priced higher, the completion curve is usually steeper — but the result depends on your specific ratio. The analyzer reports the cost at the low and high end of each sweep and the swing each lever produces, so the dominant factor is obvious.

Tips for acting on the result

If output dominates: set a tighter max_tokens, ask for concise answers, and avoid prompting for verbose explanations you don’t read.
If input dominates: compress the prompt, dedupe context, and use prompt caching so stable context bills at a fraction of the input rate.
Re-check after model changes. Switching to a model with different input/output pricing can flip which lever matters.
Optimize the dominant lever first. A 20% cut on the lever that drives 80% of cost beats a 50% cut on the one that drives 20%.