Claude pricing seems complicated until you understand 3 concepts: tokens, cache, batch. This guide gives you everything to budget and optimize your bill.
What is a token
A token is approximately 4 characters or 3/4 of an English word. "Hello world" is ~3 tokens. 1000 tokens ≈ 750 words.
Anthropic charges separately for input (what you send) and output (what it responds). Output usually costs 3-5x more than input.
Price structure
| Model | Input | Output |
|---|---|---|
| Haiku | ~$0.80 | ~$4 |
| Sonnet | ~$3 | ~$15 |
| Opus | ~$15 | ~$75 |
Real cost examples
- Summarize 1000-word article with Sonnet: ~$0.006.
- Process 10k short emails with Haiku: ~$8-15 total.
- Full refactor of 500-line file with Opus: ~$0.30-1.00.
Prompt caching: the big saver
If you send same long context every request (a manual, CLAUDE.md, examples), activate prompt caching. Save up to 90% on repeated input cost.
Batch API
For non-urgent tasks, send batch: Anthropic processes within 24h at half price.
Monthly budget calculation
- Estimate requests per month.
- Average input and output tokens per request.
- Multiply by model price.
- Add 30% margin.
Bill-cutting tricks
- Right model.
- Cache repeated context.
- Batch API for non-urgent.
- Limit max_tokens.
- Compress prompts.
- Deduplicate.
Rate limits
Each tier has TPM and RPM limits. If exceeded, API returns 429. Handle with exponential backoff.
Conclusion
Claude API cost is predictable if you follow 3 rules: right model per task, caching when applicable, batch for non-urgent. Many small apps live well with $50-100/mo.