What is a token
The unit you are billed in. Getting a feel for it is the difference between a $50 and a $500 bill.
Rough conversions
| Content | Tokens |
|---|---|
| 1 English word | ~1.3 |
| 1 Chinese character | ~0.6–1.0 |
| 1 page of English text (500 words) | ~650 |
| 1 page of Chinese text | ~500–700 |
| 1 line of code | ~10–15 |
| A 300-page book | ~120,000 |
Chinese and English do not tokenise the same way. Chinese characters are handled in clusters, so the ratio is roughly one token per 1–1.7 characters depending on the text.
Why it matters in practice
| Scenario | Input tokens | Output tokens |
|---|---|---|
| Short chat turn | 200 | 300 |
| Document summarisation (10 pages) | 7,000 | 500 |
| Code review of one file | 3,000 | 1,500 |
| RAG answer with 5 retrieved chunks | 4,000 | 600 |
| Agent task, 20 turns | 60,000+ | 15,000 |
The three multipliers that surprise people
1. Output costs more than input
Flash charges $0.15 per million input but $0.60 per million output — four times more. V4 Pro is three times more. If you generate long responses, output dominates the bill.
2. Every turn re-sends the history
A chat with 10 turns does not cost 10 times one turn; it costs about 55 times, because each turn includes all previous turns. This is why agents escalate so quickly.
3. Cached input is almost free
A cache hit costs 1/50th of a miss. Stable content at the front of the request gets cached.
Estimate yours
Paste your text into the token estimator, or use the cost calculator with your daily volume.
Related
Cost calculator · Pricing · Off-peak guide · Token reference