Prompt caching by provider
The mechanism is the same everywhere — an exact-prefix match, cached reads at a fraction of base price. The minimums, the pricing and the failure modes are not.
| Provider | Min tokens | Read | Write | Base $/MTok |
|---|---|---|---|---|
| Anthropic Claude | 1,024 | 0.1× | 1.25× | $3.00 |
| OpenAI | 1,024 | 0.1× | 1.25× | $5.00 |
| Google Gemini | 2,048 | 0.1× | 1× | $1.25 |
| DeepSeek | 1,024 | 0.1× | 1× | $0.28 |
The minimum is often per model, not per provider. On Anthropic it ranges from 512 to 4,096 tokens and runs backwards from price — see the Anthropic page for the full table.