Bedrock vs Anthropic API: 2026 Claude Cost Guide

Bedrock vs Anthropic API comes down to this: on global endpoints, Claude costs the same per token on Amazon Bedrock and on Anthropic’s own Claude API. The real gap shows up in four places: the 10% regional premium, batch discounts that Bedrock’s new Claude integration does not offer, missing server-side features, and which company bills you and processes your data. This guide puts the official numbers side by side and runs one monthly workload through each path.

Checked: November 10, 2026 (KST). Prices are list prices in USD from Anthropic’s pricing page, Anthropic’s Bedrock documentation, and AWS pricing pages as of that date. Private offers and negotiated discounts can change them.

Bedrock vs Anthropic API: the short answer

At global endpoints, Claude’s per-token list price is identical on Amazon Bedrock and on the Anthropic API. Bedrock costs more only when you pin traffic to a region (10% premium) or need batch processing, which Bedrock’s new Claude integration does not support. Anthropic’s API costs more in AWS-specific ways: it is billed outside AWS and does not retire an AWS spend commitment.

Anthropic’s own documentation now calls its first-party service the Claude API, and it states the rule directly: regional and multi-region endpoints on Bedrock and Google Cloud “include a 10 percent premium over global endpoints,” while the Claude API “is global by default.” That leaves token price as a tie for most teams that run on global routing.

So the useful question is not “which one has cheaper tokens?” It is “which line items does my workload actually trigger?” Those line items are:

• Routing: global (no premium) or pinned to a geography (10% more on Bedrock, 1.1x on the Claude API with US-only inference).
• Batch: 50% off on the Claude API’s Message Batches. Not available on Bedrock’s new Messages API integration for Claude.
• Caching: the same multipliers on both, but cache behavior can vary by Bedrock region.
• Features: server-side tools like web search and code execution exist only on the Anthropic-operated side.
• Billing and data: AWS bills and processes data on Bedrock. Anthropic does both on the Claude API.

Token prices side by side

For current models on global endpoints, input, output and cache prices match across the Claude API and Bedrock. Opus 5.5 is $4 input and $20 output per million tokens on both. Sonnet 5.5 and Sonnet 5 are $2 and $10. Haiku 5.5 starts at $0.10 and $0.50.

The table below uses Anthropic’s published rates. Anthropic’s Bedrock page says the only pricing difference it lists for Bedrock is the 10% regional premium, so the global column applies to both platforms. “MTok” means one million tokens.

How to read this table: the regional column is a 10% calculation on Anthropic’s list price, not a separate AWS quote. AWS’s own Bedrock pricing page did not show per-model Claude tables when we checked, so confirm the final number in the AWS Pricing Calculator for your region before you budget.

Two details change real bills more than the headline rates:

• Tokenizer change. Anthropic says Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text.” If you compare an older model on one platform with a newer one on the other, you are not comparing equal token counts.
• Haiku 5.5 length tiers. Prompts over 100,000 tokens jump to $0.50 input and $2.50 output, and the whole request is priced at the higher tier.

The 10% regional and US-only premium

Bedrock charges 10% more when you use a regional endpoint or a geographic inference profile (US, EU, JP, AU) instead of the global endpoint. The Claude API has a similar 1.1x multiplier when you set inference_geo to “us.” Either way, data residency costs about the same.

On Bedrock’s new Claude integration (Opus 4.7 and later, plus Haiku 4.5), Anthropic’s documentation describes two endpoint types:

• Global: dynamic routing across all available regions, with no pricing premium.
• Regional: resolves to one region for data residency, with a 10% premium over global. Geographic inference profiles keep traffic inside the US, EU, Japan or Australia.

On the Claude API, global routing is the default. For Claude 4.6 and later models, sending inference_geo: "us" applies a 1.1x multiplier to every token category, including cache reads and writes. Anthropic notes that this multiplier stacks with batch and caching discounts.

The takeaway: if your compliance team requires US-only processing, plan for roughly 10% on either platform. The premium alone is not a reason to pick one side.

Batch and prompt caching: where the bill really splits

Batch is the biggest price gap. The Claude API’s Message Batches cut input and output prices by 50% for current models. Anthropic’s Bedrock documentation lists Message Batches as not supported on the new Claude integration, and AWS’s batch inference list does not include Opus 5.5, Sonnet 5 or Sonnet 5.5.

Anthropic’s batch table shows Sonnet 5.5 at $1 input and $5 output, and Opus 5.5 at $2 and $10. If a job can wait for asynchronous results, such as nightly classification, document tagging or eval runs, that is half the token bill.

On Bedrock, AWS runs its own batch inference feature at 50% of on-demand pricing for supported models. As of November 10, 2026, AWS’s supported-models page lists Claude models such as Haiku 4.5, Sonnet 4.6, Opus 4.6 and Opus 5, but not Opus 5.5, Sonnet 5 or Sonnet 5.5. So on Bedrock you can batch, but you may have to use an older or different model to do it.

Prompt caching is closer to even. Anthropic’s multipliers are:

• 5-minute cache write: 1.25x the base input price.
• 1-hour cache write: 2x the base input price.
• Cache read: 0.1x for most models, 0.05x for Opus 5.5 and Sonnet 5.5, and 0.025x for Fable 5.2 and Fable 5.1.

Bedrock’s new integration supports prompt caching. Third-party cost guides report that cache support can vary by Bedrock region, so check your CloudWatch cache metrics in the first week. A cache count of zero means you are paying full input price.

Costs beyond tokens: network, logs, quotas and commitments

For text workloads, network and logging charges are usually small next to token spend. The costs that matter beyond tokens are missing features you would have to rebuild, Bedrock quotas, and whether the spend retires your AWS commitment.

Network path

Calling Bedrock from private subnets usually goes through an AWS PrivateLink interface endpoint. AWS bills those per endpoint per Availability Zone per hour, plus $0.01 per GB processed for the first petabyte. AWS’s own pricing example uses $0.01 per hour per endpoint network interface, which works out to about $21.90 a month for three Availability Zones.

Calling the Claude API from the same subnets usually goes out through a NAT gateway. CloudZero’s AWS egress guide cites NAT data processing at $0.045 per GB, on top of the gateway’s hourly charge. Keep in mind that cached tokens are still sent on every request. Caching lowers the bill, not the bytes on the wire.

Logging

Bedrock model invocation logging is off by default. When enabled, it can write full request and response bodies to CloudWatch Logs or S3, with bodies over 100 KB sent to S3. AWS’s pricing examples use $0.50 per GB for standard CloudWatch Logs ingestion and $0.03 per GB for archive, with 5 GB free each month. One catch from AWS’s documentation: invocation logging captures bedrock-runtime calls, and calls through other endpoints such as bedrock-mantle “are not currently captured.” The new Claude integration uses the bedrock-mantle endpoint, though Anthropic notes that activity is still emitted to CloudWatch and CloudTrail.

Quotas

Anthropic’s Bedrock page lists a default of 2 million input tokens per minute. You can request up to 5 million input TPM and 500,000 output TPM without extra Anthropic approval. Requests-per-minute limits are handled by AWS Support.

Feature gaps you pay for in engineering time

Bedrock’s new Claude integration does not support the Files API, URL sources for images and documents, web search, web fetch, code execution, Agent Skills, the MCP connector, Message Batches or Claude Managed Agents. If your app needs web search, on Bedrock you build and pay for it yourself.

Commitments and billing

Bedrock is a native AWS service, so it lands on your AWS bill. The direct Claude API is invoiced by Anthropic. If your company has an AWS spend commitment, that difference can matter more than any token price. Ask your AWS account team how Bedrock and Marketplace spend count under your specific agreement.

A worked monthly example

In this example workload on Sonnet 5.5, the Claude API with batch costs about $1,278 a month. The same work on Bedrock’s global endpoint costs about $1,678, and about $1,846 on a regional endpoint. The gap comes almost entirely from batch, not token prices.

This is a hypothetical workload, not a measured bill. It uses list prices for Claude Sonnet 5.5 and these inputs:

• Real-time chat: 100M uncached input tokens, 20M tokens written to the 5-minute cache, 280M cache-read tokens, 60M output tokens.
• Nightly job: 200M input tokens and 40M output tokens that can wait for asynchronous results.

Formula: tokens in millions × price per MTok. The regional column multiplies each line by 1.1. Network and logging are left out because, at these volumes, they are a few dollars to a few tens of dollars a month.

What changes the answer:

• No batchable work: the totals become $878 on both global options. Price stops being a factor.
• US-only requirement: the real-time part is $965.80 on Bedrock regional and on the Claude API with inference_geo: "us".
• Batch on Bedrock with another model: you can move the nightly job to a model on AWS’s batch list. Remember that Sonnet 4.6 uses the older tokenizer and has a higher list price ($3 / $15), so re-run your own token counts before comparing.

The third path: Claude Platform on AWS

Claude Platform on AWS, generally available since May 11, 2026, gives you Anthropic’s own platform through your AWS account. It bills through AWS Marketplace at standard Claude API rates and supports batch processing, but Anthropic, not AWS, processes the inference data.

This option sits between the other two. According to Anthropic’s documentation:

• Billing: usage is converted to Claude Consumption Units (CCUs) at $0.01 per CCU, metered hourly and invoiced monthly on your AWS bill. CCUs are not prepaid credits.
• Features: most Claude API features, including batch processing, prompt caching, the Files API and Managed Agents. Fast mode, OAuth and the HIPAA-ready program are not supported.
• Commitments: Anthropic’s migration notes list “AWS commitment retirement” under “What stays the same” when moving from Bedrock.
• Data: Anthropic is the data processor for inference inputs and outputs. Zero Data Retention is opt-in through your Anthropic account representative.
• Discounts: negotiated discounts and Bedrock private offers do not transfer automatically.

If you want batch pricing and Anthropic’s full feature set but your finance team wants one AWS invoice, this is the path to price out. If you need AWS to be the only data processor, or need FedRAMP High or HIPAA readiness, Anthropic points you back to Bedrock.

How to choose, step by step

Start with compliance, then batch, then features, then billing. If AWS must be the sole data processor, use Bedrock. If a large share of your tokens can be batched on current models, the Claude API or Claude Platform on AWS will usually cost less.

A five-step check before you commit:

1. Export one month of token usage split into uncached input, cache writes, cache reads and output.
2. Tag which requests can tolerate batch latency.
3. Confirm with legal whether you need regional routing or a specific data processor.
4. List the server-side features your app uses today.
5. Ask your AWS account team how Bedrock and Marketplace spend count toward your agreement.