System Prompt Token Cost
What your system prompt costs per request, per million requests and per month, at your own rates.
~105 estimated tokens × your $3 rate. A million requests at a per-million rate cancels out, so the dollar figure is simply the token count times the rate.
- System prompt
- 441 chars · ~105 tokens (est.)
- Per request
- $0.00032
- Per 1,000 requests
- $0.32
- Per 1,000,000 requests
- $316.40
- At 500,000 requests/month
- $158.20 per month
- Blended rate at a 0.0% hit rate
- $3.7500 per 1M tokens
- Monthly with caching
- $197.75
- Monthly difference
- costs $39.55
- Each 100 tokens you delete
- $150.00 per month
usage object on an actual API response: that is what you are billed on, and it includes chat-template wrapping this model cannot see. The system prompt is assumed to be sent on every request, which is what makes it worth measuring — it is the only part of a prompt you pay for a million times. Misses are billed at the cache write rate and hits at the cached read rate, so a hit rate of zero with a write premium correctly shows caching costing you money rather than saving it. Output tokens are not modelled here at all: this is the input side only, and output is normally the more expensive one.Why a system prompt is the one text worth auditing
Everything else in a request varies. The system prompt does not — it is the same tokens, on every call, forever, which turns a paragraph you wrote once into a recurring line item. The identity in the headline is the reason to care: at a rate quoted per million tokens, the cost of one million requests is just the token count multiplied by the rate. A 400-token system prompt at $3 per million is $1,200 per million requests. Add three examples and a formatting section and take it to 1,400 tokens, and it is $4,200 for the same traffic.
The last row is the actionable one. It prices 100 tokens of deletion at your own volume, and it is usually the number that decides whether an afternoon spent tightening the prompt pays for itself. At small volumes it will not, and the honest answer is to leave the prompt alone.
Two caveats worth stating. First, prompt caching changes the arithmetic more than editing does, because a stable prefix read from cache is typically billed at a fraction of the base rate — but only if it is genuinely stable and long enough to qualify, which is a question about your prefix rather than your prompt’s length. Second, a shorter system prompt is not automatically cheaper overall. Removing the instruction that keeps answers to three sentences saves 12 input tokens and can add hundreds of output tokens per call, at a rate that is usually several times higher. Measure the total, not the half you can see.