AI cost · AWS

Amazon Bedrock cost: what you actually pay

Bedrock itself adds no separate markup line — cost hides in which model you default to (a 140x spread from Nova Micro to Claude Opus), unmetered Provisioned Throughput commitments, and the Knowledge Bases/Guardrails/Agents charges that stack on top of raw tokens.

Public pricing

PlanPriceBillingIncludes
On-DemandCustomNo commitment; pay per input/output token at the invoked model's published rate.usageAccess to all on-demand Bedrock foundation models
Batch InferenceCustomAsync jobs submitted via S3 input/output for supported models.usageSame models as on-demand at roughly half the per-token rate
Provisioned ThroughputCustom1-month or 6-month commitment; custom quote from your AWS account team; required for guaranteed throughput or some custom/fine-tuned models.customDedicated capacity with no shared-tenancy throttling

Token rates (USD per 1M tokens)

ModelInputOutputCache read
Claude Opus 4.7 (Anthropic)On-demand; matches direct Anthropic API rate.$5$25$0.5
Claude Sonnet 4.6 (Anthropic)On-demand; batch inference runs ~50% off ($1.50/$7.50).$3$15$0.3
Amazon Nova ProAWS's own flagship multimodal model; Nova Micro (text-only) runs as low as $0.035/$0.14 input/output.$0.8$3.2

Bedrock charges the same per-token rate as calling most providers directly (Claude on Bedrock = Claude direct API price); Amazon's own Nova family is the cheapest in-catalog option. Provisioned Throughput needs a custom AWS quote and bills for the full commitment term regardless of utilization. Bedrock token spend also stacks with the AWS services built around it — Knowledge Bases (OpenSearch/vector store + embedding calls), Guardrails, Agents, and Data Automation all bill separately from the LLM token line.

Approximate — verify before relying on it. Last updated: 2026-07-22. Sourced from Amazon Bedrock Pricing (official), Anthropic Claude API Pricing (official).

Where the real Amazon Bedrock cost hides

The sticker is the smallest part of the story. For Amazon Bedrock, these are the line items that quietly inflate the bill:

  • Output tokens run 4-6x input tokens across nearly every model family — verbose responses and chain-of-thought-style output quietly dominate the bill.
  • The model catalog spans a ~140x price range (Nova Micro ~$0.035/MTok input vs. Claude Opus $5/MTok) — defaulting every task to a flagship model is the single biggest waste driver.
  • Knowledge Bases, Guardrails, Agents, and Data Automation bill separately from tokens — a RAG pipeline's real cost is tokens + vector store + embeddings + guardrail checks, not just the model line.
  • Provisioned Throughput is billed hourly for the full commitment term whether or not traffic shows up, and its price isn't listed next to on-demand rates, so teams under- or over-commit.
  • Cross-region inference profiles and retry logic can silently double-bill tokens if failed requests aren't deduplicated before resubmission.

What an audit finds on Amazon Bedrock

Teams default every call to a large flagship model (Opus/Sonnet) instead of routing simple extraction/classification work to Nova Micro/Lite or Haiku, skip prompt caching, and let verbose prompts run — token bills land 5-10x higher than a right-sized, cached routing setup.

Amazon Bedrock pricing FAQ

Does Bedrock cost more than calling Anthropic or Meta directly?

No — Bedrock's on-demand token price matches the model provider's direct API rate; any premium shows up indirectly via AWS networking, storage, and the surrounding services, not the token rate itself.

When is Provisioned Throughput worth it?

Only at sustained, predictable high volume — it's billed hourly for the full commitment regardless of usage, so bursty or low-volume workloads are cheaper on-demand.

Is prompt caching available on Bedrock?

Yes for supported Claude models — cache reads run about 90% cheaper than base input tokens, while cache writes cost a premium, so it only pays off with genuinely repeated/shared context.

What is Amazon Bedrock really costing you?

Spendassay connects your Amazon Bedrock spend to actual usage and engineering output — a proof-level AI cost report with the wasted dollars named. Neutral across every AI vendor.

Free · read-only · no card

More: the complete AI cost management guide · every tool, with prices · cost per outcome · compare platforms · how we measure