If you’ve ever looked at an AI pricing page, you’ve seen two prices for the same model: one for input tokens and one for output tokens. They’re not close. Output usually costs several times more.
Both are Tokens, the little chunks of text AI reads and writes. The difference is just which direction they’re going.
Input Tokens
Input tokens are everything you send the model. Your question, the instructions behind it, any document you pasted in and the whole conversation so far. In a chat app, that whole history gets sent again every time you hit enter, which is a big reason long conversations cost more to run.
They’re the cheap side. Anthropic’s Claude Sonnet 5.5 charges $2 per million input tokens, and OpenAI’s GPT-6.1 Sol charges about the same. That’s roughly a few novels’ worth of reading for the price of a coffee. A big reason it’s cheap: the model can process all your input at once, in parallel, instead of one piece at a time.
Output Tokens
Output tokens are everything the model writes back. The answer, the summary, the code. Those same two models charge $10 per million output tokens, five times the input price. The 5x gap holds across most of Anthropic’s and OpenAI’s current lineups.
Writing is slower work for the model. It produces one token at a time, and each new one depends on all the ones before it, so it can’t do them all at once the way it reads.
There’s a sneaky extra here. When a reasoning model “thinks” before answering, that hidden thinking usually gets billed as output, too. OpenAI’s docs say reasoning tokens “are billed as output tokens,” and you never see most of them.
The tell
Ask who wrote it. If you (or your app) sent it, it’s input. If the AI produced it, it’s output, and it’s the pricier side.
It’s like a lawyer who charges a little to read your contract and a lot to write you a new one. Reading takes some time. Writing takes their full attention, one clause at a time.
For most people on a monthly plan, none of this shows up on a bill directly (it’s baked into the usage limits instead). But it explains a lot, like why “keep it short” in a prompt saves money and why turning on extra thinking costs more even when the final answer is the same length.

