AI providers publish prices per million tokens, which is a unit almost nobody thinks in. The practical questions are different: what will this feature cost me a month, is the expensive model worth it for this job, and did the price change since I last budgeted? This is the part of CalcPerch we maintain most actively — prices are re-checked weekly and every page shows the date it was last verified.
Which one do you want?
| If you are… | Use |
|---|---|
| "What will my monthly bill be?" | AI API Cost Calculator — your own volumes across every tracked model. |
| "What does this specific job cost?" | Cost Benchmarks — eight real tasks priced on every model, with the token assumptions shown so you can check the maths. |
| "How many tokens is my text?" | Token Calculator — paste it and see the count and cost. |
| "Is ChatGPT Plus worth it, or should I use the API?" | Subscription vs API — find your break-even point. |
| "Did anything change?" | Price Tracker — current list and cached-input rates plus a changelog. |
Prices come from the official pricing pages of OpenAI, Anthropic and Google. One number worth internalising: output tokens cost three to six times more than input, so anything that writes a lot costs far more than the headline rate suggests.
What makes an AI bill different from the price list
- Output costs four to eight times what input costs. Across the models tracked here the multiplier runs from 4.00× to 8.33×. A workload that writes a lot is priced by a different number than the one in the headline, so cheap-on-input and cheap-for-us are separate questions.
- Chat re-sends the whole conversation on every turn. A thousand eight-turn conversations move about 12 million input tokens, not the 1.5 million a single-message test suggests. Estimates built from one-shot testing come in low by roughly that factor.
- Prompt caching often beats dropping a model tier. Re-sent history is exactly the cacheable part, and cached input bills at roughly a tenth of the input rate — frequently a larger saving than a full step down the ladder, and it costs no capability.
- Some models are priced identically to each other. Six of the sixteen tracked here share a price with another model. Where that is true there is no cost argument to be had, and the decision has to be made on capability, latency or rate limits.
Frequently asked questions
How often is the pricing updated?
Weekly, and immediately after any major provider announcement. The "last verified" date at the top of each page is the honest freshness marker.
Do these include caching and batch discounts?
The tables show standard list pricing, which is the ceiling. Prompt caching can cut repeated input by around 90% and batch processing roughly halves eligible work — the price tracker works through both with examples.
Why is my chatbot so much more expensive than I estimated?
Because most APIs are stateless, the whole conversation is re-sent on every turn. An eight-turn conversation costs far more than eight single messages — this is the single most common budgeting mistake.