Bedrock Bill-Shock Calculator

Your real Amazon Bedrock bill, not the pricing-page guess. Every call also re-sends the system prompt, the chat history, retrieved documents and tool results, and Guardrails, the vector database and logging are billed on top. Add it all up before your first invoice does.

Start from:

Traffic Daily active users Conversations per user per day Messages per conversation

Prompt System prompt + tool definitions (tokens) Sent with every single call. 1 token is about 4 characters of English. User message (tokens) Answer length (tokens) Output costs 4-5x more than input on most models. Chat history sent with each message Messages kept

Knowledge base (RAG) Chunks retrieved per message 0 = no retrieval. Tokens per chunk Vector store Vector store cost (USD per month) S3 Vectors: leave 0 to use a $5 starting estimate. Documents to index (tokens) Re-indexing runs per month

Agents and reliability Tool calls per message (agent steps) Each step is another full model call with the whole context. Tool result size (tokens) Tokens the model writes per tool call Calls retried (%) Timeouts, throttling, invalid JSON: each retry is billed again.

Model and pricing Model

Endpoint Prompt caching for the system prompt

Traffic sent as batch jobs (%) Batch is 50% cheaper for work that can wait.

Use your own prices for this model

Input (USD per 1M tokens) Output (USD per 1M tokens)

Guardrails and logging Content filters Denied topics Sensitive information (PII) filter Contextual grounding check What Guardrails checks Model invocation logging to CloudWatch Logs

Why Bedrock bills surprise teams

History is re-sent

Message 10 of a chat sends the previous 9 questions and answers again. Input grows with every turn.

Context is re-sent

The system prompt, tool definitions and retrieved documents ride along on every call, often 10-50x the question itself.

Agents multiply calls

One question can become 3-10 model calls when an agent uses tools, each one with the full context.

The extras are billed separately

Guardrails per 1,000 characters per policy, an OpenSearch Serverless minimum around the clock, and logging per GB.

Guardrails I set up so this never becomes a surprise

AWS Budgets and Cost Anomaly Detection alerts on the Bedrock spend
A hard daily and monthly cap with a kill switch, per user and per feature
max_tokens on every call, a history window, and prompt caching
Cost per feature with application inference profiles and tags
Guardrails scoped to user input, with PII redaction before logging

Want your setup checked? The GenAI Cost & Safety Check is a fixed-price review with a written report and the fixes.

See the service