Output tokens
Also called Completion tokens
What the model writes: replies, code, tool calls, hidden thinking. Produced one by one, so they set the wait, and they cost the most.
Output tokens are what the model writes back on each model provider request: the visible reply, the code it produces, the structured tool call it emits, and any hidden reasoning when effort is turned up. They're generated one at a time by next-token prediction, which makes them slow to produce, and providers price them several times higher than input tokens to match.
Most sessions spend far more on input than output, so output tokens are rarely the cost problem. They are the latency problem. A model rewriting a 600-line file from scratch is emitting thousands of tokens you then wait for, when a targeted diff would have been fifty. Extended thinking has the same shape: minutes of reasoning tokens on a task that didn't need any.
Ask for edits, not rewrites, and let the harness apply diffs. Keep reasoning effort proportional to the difficulty. When you want a long artefact (a spec, a migration plan), that's a fine use of output tokens; when you want a one-line change, a long answer is a smell, not a bonus.
- Claude CodeFile edits are applied as diffs, so output stays small; thinking tokens show up under output in
/cost. - Anthropic API
max_tokenscaps output per request; thinking, if enabled, counts against it.
“It regenerated the entire component to change one className.”
“That's a few thousand output tokens for a five-token change. Tell it to edit in place.”
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →