OpenAI API

What OpenAI actually meters

OpenAI meters tokens, and it splits them three ways rather than two. Input, cached input and output each carry their own rate, and the gaps between them are large enough to change how you build.

Output is the expensive half. GPT-6 Astra reads at $10 per million and writes at $50, a 5x spread. GPT-5.6-Terra reads at $2 and writes at $12, a 6x spread. A verbose system prompt is cheap by comparison with a verbose answer, so instructing a model to be concise is a pricing decision as much as a product one.

Cached input is where the real discount sits. Most current models price cached input at 90% below standard input, so GPT-5.6-Sol drops from $4 to $0.40 per million. That only pays off if the stable part of your prompt sits at the front and stays byte-identical between calls, which makes prompt ordering a billing concern.

GPT-4o is the exception worth noting. Its cached input is $1.25 against $2.50 standard, a 50% discount rather than 90%, so the older model does not benefit nearly as much from caching.

What happens when you hit the limit

Nothing to hit. There is no included allowance on the API, so there is no overage concept. Every token bills at list rate from the first call.

Spend control therefore lives entirely in account-level configuration rather than in the pricing model: prepaid credit balances, usage limits and rate limits. The pricing page does not document credit expiry, auto top-up behaviour or what happens when a prepaid balance reaches zero, so those mechanics need verifying against account settings before publish.