Prompt Cost Analyzer
Estimate the monthly cost of a prompt across OpenAI's GPT-5 models — and see how prompt caching changes the math at different traffic patterns.
Cost shape
Cache only discounts the system sliceSystem is only 3.7% of cost — caching can't move the bill much in this shape. Either increase the system prompt or decrease output tokens to see meaningful cache savings.
Monthly cost
Pricing snapshot: 2026-05-12. Estimates use the token counts and monthly calls above.
- No cache · 0% hit rate
- A changing prefix or a prompt below the cache threshold.
- Bursty · 50% hit rate
- Intermittent calls, with gaps between requests.
- Steady · 80% hit rate
- Regular traffic, with occasional gaps.
- Always-warm · 95% hit rate
- Sustained traffic with an unchanged prefix.
These hit rates are comparison assumptions, not measured performance or guarantees.
Scroll the table horizontally to compare every scenario.
| Model | Tokens In / Out | $ / Call | No cache | Bursty | Steady | Always-warm |
|---|---|---|---|---|---|---|
GPT-5.5 in $5 · out $30 · cache $0.50 | 60 → 200 | $0.0063 | $630 | $619 −1.7% | $613 −2.7% | $610 −3.2% |
GPT-5.4 in $2.50 · out $15 · cache $0.25 | 60 → 200 | $0.00315 | $315 | $310 −1.7% | $307 −2.7% | $305 −3.2% |
GPT-5.4 mini in $0.75 · out $4.50 · cache $0.075 | 60 → 200 | $0.000945 | $94.50 | $92.91 −1.7% | $91.96 −2.7% | $91.49 −3.2% |
GPT-5.4 nano in $0.20 · out $1.25 · cache $0.02 | 60 → 200 | $0.000262 | $26.20 | $25.78 −1.6% | $25.52 −2.6% | $25.40 −3.1% |
How caching works here
This is a single-turn model — system prompt is treated as the cacheable prefix; user prompt and output tokens bill at standard rates. In a multi-turn conversation or agent loop, the cacheable prefix actually grows to include prior turns, so cache savings scale further than what's shown here. On a cache hit, prefix tokens bill at OpenAI's discounted cache-read rate — about 10% of standard input pricing for the GPT-5 family. Real applications with long, stable prefixes and high call volume often sustain 80%+ hit rates, turning the cache columns into the actual P&L line.
Token counts are approximate (±10%). Estimated from character length; actual counts will differ slightly by content.
Source · verified 2026-05-12
OpenAI API pricing (opens in new tab)Need help optimizing your AI bill?
I run Devclock. Get in touch if you'd like a second pair of eyes on your prompt structure, model selection, or cost trajectory.
Visit DevclockStay in the loop
The occasional email when there's something new worth sharing.