Bedrock models with prompt caching
These Amazon Bedrock models support prompt caching: a prefix you send repeatedly is stored server-side and re-read at a fraction of the input price. For anything that resends a long system prompt — agents, RAG with a fixed instruction block, chat with history — this is the single largest lever on the bill.
21 of the models on Bedrock qualify today, from $0.035 per 1M input tokens. Rebuilt daily from the AWS catalogue, so this list does not go stale.
What teams get wrong about this
Caching and batch inference do not combine on Bedrock. You pick one. Batch halves the rate for work that can wait; caching cuts the resent prefix for work that repeats. Choosing the wrong one for the workload is a common and expensive mistake, and the discount column below only applies to the cached portion of your input, never to output.
All 21 models, cheapest first
| Model | Input /1M | Cache read |
|---|---|---|
| Nova Micro amazon.nova-micro-v1:0 | $0.035 | 4× less |
| Nova Lite amazon.nova-lite-v1:0 | $0.060 | 4× less |
| GPT-5.6 Luna openai.gpt-5.6-luna | $0.200 | 10× less |
| Nova 2 Lite amazon.nova-2-lite-v1:0 | $0.300 | 4× less |
| Nova Pro amazon.nova-pro-v1:0 | $0.800 | 4× less |
| Claude Haiku 4.5 anthropic.claude-haiku-4-5-20251001-v1:0 | $1.00 | 10× less |
| GPT-5.6 Terra openai.gpt-5.6-terra | $2.00 | 10× less |
| Grok 4.6 xai.grok-4.6 | $2.00 | 4× less |
| Nova Premier amazon.nova-premier-v1:0 | $2.50 | 4× less |
| Claude Sonnet 4.5 anthropic.claude-sonnet-4-5-20250929-v1:0 | $3.00 | 10× less |
| Claude Sonnet 4.6 anthropic.claude-sonnet-4-6 | $3.00 | 10× less |
| Claude Sonnet 5 anthropic.claude-sonnet-5 | $3.00 | 10× less |
| GPT-5.6 Sol openai.gpt-5.6-sol | $4.00 | 10× less |
| Claude Opus 4.5 anthropic.claude-opus-4-5-20251101-v1:0 | $5.00 | 10× less |
| Claude Opus 4.6 anthropic.claude-opus-4-6-v1 | $5.00 | 10× less |
| Claude Opus 4.7 anthropic.claude-opus-4-7 | $5.00 | 10× less |
| Claude Opus 4.8 anthropic.claude-opus-4-8 | $5.00 | 10× less |
| Claude Opus 5 anthropic.claude-opus-5 | $5.00 | 10× less |
| Claude Fable 5 anthropic.claude-fable-5 | $10.00 | 10× less |
| Claude Opus 4.1 anthropic.claude-opus-4-1-20250805-v1:0 | — | — |
| Claude Sonnet 4 anthropic.claude-sonnet-4-20250514-v1:0 | — | — |
Capability flags come from the AWS Bedrock model catalogue; prices from the Price List API. Region counts include cross-region inference profiles. Methodology.
Questions
How much does prompt caching save on Bedrock?
A cache read is typically billed at about one tenth of the normal input rate, and writing to the cache costs slightly more than a normal input token. The saving therefore depends entirely on your hit rate: a prefix reused across many calls pays the write once and the cheap read many times, while a prefix used twice barely breaks even.
Can I use prompt caching and batch inference together?
No. On Bedrock these are mutually exclusive. Batch is the right choice for large offline jobs with no repeated prefix; caching is the right choice for online workloads that resend the same context.
Does caching apply to output tokens?
No. Only input. Output is always billed at the full rate, which is why output-heavy workloads see less benefit than the headline discount suggests.