AI Tokenomics in a Minute
Tokens
What tokens are in LLM APIs, how input and output differ, and why token counts drive coding-agent cost.
In one sentence
A token is a model’s billing and context unit for pieces of text (and sometimes other modalities)—not exactly a word, and not free.
Why it matters
Coding agents spend tokens on prompts, files, tool results, and completions. Misunderstanding tokens makes cost and context-limit surprises look like mysteries.
How it works
- Providers tokenize text into subword pieces; counts vary by model family.
- Input tokens cover what you send (system prompts, files, tool output).
- Output tokens cover what the model generates.
- Some products expose cached, reasoning, or other token classes with different prices.
- Context windows limit how many tokens can be present at once; growth raises both cost and truncation risk.
Example
A short English sentence might be ~10–20 tokens depending on the tokenizer. A 2,000-line source file pasted into context can be thousands of tokens—and may be re-sent on every turn unless caching or compaction applies.
What this proves
Token counts from a provider prove how that provider billed or measured a request under its tokenizer and product rules.
What this does not prove
Token spend does not prove task success, code quality, or that a cheaper model would have failed.
Last reviewed 2026-09-06. Title for citation: AI Tokenomics in a Minute: Tokens.