Both matter for cost, speed, limits, and how much room the model has to work.
Think about an internet connection. Download is what comes into your house; upload is what you send out. They’re related, but not the same lane, and one can be slower, more limited, or priced differently. Tokens work the same way. The incoming side is your prompt, the conversation so far, files, transcripts, instructions, and tool results. The outgoing side is the answer, draft, or code the model writes back. It’s easy to assume only your prompt counts, but a long session history, a giant attached document, and a tool result the agent reads are all input too.
How it shows up
Ask an AI to write a client memo from three transcripts: the transcripts and your instructions are input tokens, and the memo is output tokens. Say “make it shorter” and the model rereads the prior memo as input, then writes a new version as output, which is why long sessions get expensive and crowded. Output also behaves differently. A model can usually read far more than it can write at once: your context window might hold a huge amount of input, but the maximum answer is much smaller, and output is slower because the model generates one chunk at a time. This is why prompt caching helps when the same instructions or documents get reused across runs.
Why you care
Input and output explain a lot of AI friction. Slow session? Too much input. Answer stopped early? An output limit. Bill jumped? You had it read and write more than you realized. You don’t need to count tokens by hand, just carry the instinct: give the model the context it needs, not a pile of maybe-relevant material, and ask for the output you actually want.