How Amazon Bedrock pricing works
For text models, you pay per token: a price per million input tokens (everything you send, including the system prompt, documents and conversation history) and a usually higher price per million output tokens (what the model writes). Four choices change that price a lot. This calculator lets you set each of them.
1. The model
Prices range from a few cents per million tokens for small open models to tens of dollars for frontier models. Use Compare models for this workload to rank every model by the cost of your own workload. Then test the cheapest few on your real prompts, because quality and token usage differ.
2. Region and endpoint
Prices and availability vary by region. With cross-region inference, Bedrock routes requests across regions. A global profile is often cheaper than a geographic one (for example US-only). For Kimi 3, AWS lists $3.00 global vs $3.30 US per million input tokens. Some models are only offered through cross-region inference.
3. Service tier or batch
- Standard: normal on-demand.
- Flex: about 50% cheaper for supported models, but requests can take longer. Good for background jobs.
- Priority: about 75% more, for latency-critical traffic.
- Batch: 50% cheaper for supported models. You submit a file of requests and get the results later, which suits summarising yesterday's call transcripts, for example.
- Reserved / Provisioned Throughput: committed capacity, priced through your AWS account team.
4. Prompt caching
If many requests start with the same text — a long system prompt, a policy document, the conversation so far — caching can cut input cost sharply. Cache reads are much cheaper than normal input, but writes can cost more.
- Claude: writes cost about 1.25× the input price for a 5-minute cache or 2× for a 1-hour cache, and reads about 0.1×.
- Amazon Nova: reads cost 75% less, with no extra write charge.
The estimator shows how much of your input repeats, and therefore what caching could save.
Estimating tokens
Roughly 1 token is about 4 characters of English text. In a chat, each turn usually re-sends the whole history, so input grows with every turn. A 6-turn conversation can use several times the tokens of 6 single questions. The workload estimator does this sum for you, and the templates include an Amazon Connect agent assistant, a RAG chatbot, a Lex voice bot, document processing, and batch transcript summaries.
The rest of the solution
- Guardrails: each enabled policy is charged per 1,000 text units (up to 1,000 characters each):
- content filters and denied topics: $0.15;
- sensitive information filters and contextual grounding checks: $0.10;
- Automated Reasoning: $0.17 per policy;
- word and regex filters: free.
- Knowledge Bases (managed): $5 per GB of raw data per month for index storage, $1 per 1,000 retrievals, and $4 per 1,000 agentic retrievals (plus $1 per 1,000 of the Retrieve calls they make). Cohere Rerank 3.5 is $2 per 1,000 queries.
- Data Automation: $0.010 per page for standard document output, $0.040 per page for custom output (up to 30 fields), $0.006 per audio minute and $0.050 per video minute.
- Flows: $0.035 per 1,000 node transitions.
For an Amazon Connect solution, add the Connect channel costs from the Amazon Connect pricing calculator.
Keeping the prices current
Each model shows where its price came from:
- checked: compared with the AWS pricing page on 17 October 2026;
- AWS Price List: read automatically from AWS's machine-readable price list, refreshed weekly;
- not re-checked: confirm on the AWS page before relying on it.
Always confirm final numbers on the official pricing page, and set an AWS Budgets alert. The calculator can generate the command.