Claude on Amazon Bedrock vs Anthropic API Cost

Last updated 19 September 2026. This page explains claude bedrock pricing versus the first-party Anthropic Claude API: list-price parity on global on-demand endpoints, where the ~10% regional / geo premium appears, how Batch and prompt caching compare, why Fast mode stays first-party only, and when AWS governance or EDP discounts flip the net bill. Rates and partner availability evolve — confirm on Anthropic’s pricing docs and the AWS Bedrock pricing page. [VERIFY: 2026-09-19]

Not affiliated with Anthropic. Claude Token Optimization is an independent site and audit service. Claude® and related product names are trademarks of Anthropic PBC. Amazon Bedrock, AWS, and related names are trademarks of Amazon.com, Inc. or its affiliates. Plan prices, model rates, regional premiums, Batch discounts, and Marketplace SKUs change; always confirm on Anthropic’s and AWS’s own docs. [VERIFY: 2026-09-19]

Searches for claude bedrock pricing, “Bedrock vs Anthropic API cost,” and “should I run Claude on AWS” usually want the same answer: list token rates are often identical on the global path, but residency pins, partner feature lag, AWS add-ons, and enterprise discounts decide the winner. This spoke sits under the Pricing hub. Pair it with the API pricing rate card, Batch API pricing, prompt caching, cost per task, cut token usage, and how tokens work. Soft CTA to a written review at the end — no dollar price on this page.

Three paths: Bedrock, Anthropic API, and Claude Platform on AWS

Teams often collapse “Claude on AWS” into one product. There are at least three distinct billing surfaces — pick the wrong one and your invoice story will not match the docs you read. [VERIFY: 2026-09-19]

1. Amazon Bedrock (partner-operated)

Classic Amazon Bedrock InvokeModel / Converse paths: Anthropic models run as Bedrock foundation models. AWS invoices you for model usage under Bedrock SKUs. You get IAM, VPC PrivateLink, CloudTrail, CloudWatch, regional endpoints, and optional enterprise discount programs (EDPs / committed use) that can change net cost below published on-demand rates. This is the path most people mean by “Claude on Bedrock.”

2. Anthropic Claude API (first-party)

Direct Messages / Batches against Anthropic’s own API with an API key and Console. Billed by Anthropic at the published token card (see API pricing). Global by default; optional data-residency / inference_geo pins can apply a multiplier on newer models. Fastest access to new models, betas, and Fast mode.

3. Claude Platform on AWS (Anthropic-operated via Marketplace)

A different product from classic Bedrock InvokeModel: Claude Platform on AWS is Anthropic-operated via AWS Marketplace and billed in Claude Consumption Units (CCUs) at $0.01 per CCU after rating usage at standard token rates. Do not treat CCU Marketplace billing as interchangeable with Bedrock on-demand token lines when you build a forecast. [VERIFY: 2026-09-19]

The rest of this page focuses on Bedrock on-demand Claude vs first-party Claude API, with Marketplace CCUs called out only so you do not mix invoices.

List-price parity on the global path

On-demand Claude token rates on Bedrock global endpoints are generally the same USD per million tokens (MTok) as the first-party Claude API standard card. Match the live site card on API pricing: [VERIFY: 2026-09-19]

[VERIFY: 2026-09-19]

If your finance model already uses those four lines for Anthropic API spend, you can usually reuse the same unit economics for Bedrock global on-demand — before regional pins, EDPs, transfer, and add-on services. Always re-check the AWS Bedrock pricing page for the exact model ID and region you will call. [VERIFY: 2026-09-19]

Where the bill diverges: regional 10%, EDPs, transfer, add-ons

List-price parity is the starting point, not the invoice. These levers move the net dollars even when the published MTok card looks identical.

~10% regional / geo premium

Starting with Sonnet 4.5, Haiku 4.5, Opus 4.5 and later, Bedrock regional (and geo cross-region) endpoints include an approximate 10% premium over global. On the first-party API, traffic is global by default; pinning with inference_geo: "us" on Claude 4.6+ also applies roughly 1.1×. Data residency therefore costs about ~10% on both sides when you pin geography. [VERIFY: 2026-09-19]

Practical rule: if compliance does not require a US-only or regional pin, prefer the global path on whichever platform you choose. Paying +10% for residency you do not need is one of the most common silent overpays in Claude Bedrock pricing reviews.

AWS EDPs and committed use

Enterprises already deep in AWS often negotiate Enterprise Discount Programs or committed-use discounts that apply to Bedrock spend. Those can make Bedrock cheaper than list even when published MTok rates match Anthropic’s card. First-party Anthropic may have its own commercial agreements — compare your contracted nets, not only public list.

Cross-region data transfer and Bedrock add-ons

AWS can bill cross-region data transfer separately from model tokens. Bedrock add-ons — Guardrails, Knowledge Bases, AgentCore, and similar — are separate SKUs on top of Claude token lines. Provisioned throughput commitments can also change unit economics versus pure on-demand. None of these appear on the Anthropic API rate card; they are AWS-side cost levers. [VERIFY: 2026-09-19]

Batch and prompt caching on both sides

Both platforms offer meaningful volume discounts beyond raw on-demand rates — but implementation details differ.

Batch / async ~50% off

Both sides document roughly 50% off for batch / async inference: Bedrock documents 50% lower versus on-demand for select foundation models; first-party Message Batches is 50% off standard usage. Deep dive on the Anthropic path: Claude Batch API pricing. Prefer Batch when overnight latency is fine; keep sync for chat and live agents. [VERIFY: 2026-09-19]

Prompt caching multipliers

Prompt caching is available on both paths with similar multipliers for most models: roughly 5-minute write 1.25×, 1-hour write 2×, and cache hit 0.1× versus base input. Implementation differs: Bedrock typically uses explicit cache points; first-party Claude API supports automatic + explicit cache control. Stacking Batch with cache hits compounds savings when prefixes are stable — see prompt caching. [VERIFY: 2026-09-19]

Feature lag, Fast mode, and first-party-only betas

Even at identical MTok list prices, product surface is not identical.

• Feature / model lag: Newest models and features often land on the Anthropic API first; Bedrock catches up later. Plan launch calendars accordingly if you need day-one access. [VERIFY: 2026-09-19]
• Fast mode: First-party Claude API only (research preview on Opus 5 / 4.8 at roughly $10 / $50 input/output). Not on Bedrock or other partner clouds as of this verify date. [VERIFY: 2026-09-19]
• Extended-output Batches beta: Some betas (for example the output-300k Batches-only path) are first-party Batches only — not currently documented on Bedrock / other partner clouds. Treat partner availability as lagging unless AWS or Anthropic docs say otherwise. [VERIFY: 2026-09-19]

If your roadmap depends on Fast mode latency or a first-party-only beta, Bedrock list-price parity does not help — you need the Anthropic API for that slice of traffic. Many teams run a hybrid: Bedrock for steady AWS-governed volume, first-party for bleeding-edge features.

Decision matrix: when Bedrock wins vs Anthropic API

When Bedrock wins: you are already on AWS; you need IAM, VPC PrivateLink, regional residency, a single AWS invoice, a BAA via AWS, CloudWatch / CloudTrail governance, or EDP discounts that beat public list.

When Anthropic API wins: you want the fastest access to new models and features, Fast mode, a simpler key + Console path, no AWS account overhead, first-party Batch extended-output betas, or a global default without Bedrock regional markup.

Worked cost example: parity, then regional pin

Illustrative math only — re-run with your token shapes and contracted discounts. Assume 10M input + 2M output on Claude Sonnet 5 in one month, no cache hits, no Batch. [VERIFY: 2026-09-19]

At global list-price parity

• Sonnet 5 on-demand: 10M × $2 + 2M × $10 = $20 + $20 = $40 on either Bedrock global or first-party API (same card). [VERIFY: 2026-09-19]

With a ~10% regional / US pin

• Same volume × ~1.1: ≈ $44 whether you pin Bedrock regional or first-party inference_geo: "us" on eligible models. The premium is about residency, not “AWS vs Anthropic” branding. [VERIFY: 2026-09-19]

With Batch at 50% off (async-tolerant work)

• Global Batch / async at half: ≈ $20 before cache. Stacking prompt-cache hits can cut input further — see Batch API pricing and prompt caching. [VERIFY: 2026-09-19]

Now layer AWS realities: if an EDP takes 15% off Bedrock, the regional pin might still net below first-party list. If you also run Guardrails and cross-region transfer, those lines can erase the EDP win. Fold finished outcomes into cost per task rather than comparing naked MTok stickers alone.

For interactive agents that need Fast mode, the relevant comparison is not $40 vs $44 — it is whether Bedrock can serve that latency class at all. Route that slice to first-party; keep steady offline volume on whichever platform has better net governance and discounts.

Hybrid architectures teams actually ship

Many production stacks do not pick a single winner. A common pattern: route compliance-bound or VPC-only traffic through Bedrock, and keep product experiments, Fast mode paths, and day-one model launches on the Anthropic API. Shared prompt libraries and evaluation harnesses should abstract the client so you can move a workload without rewriting business logic. Meter both invoices with the same task taxonomy so “Bedrock is cheaper” claims are about finished outcomes, not different token shapes. [VERIFY: 2026-09-19]

Another hybrid: use first-party Message Batches for overnight evals that need extended-output betas, while Bedrock serves daytime sync traffic already wired into AWS IAM. The important discipline is documenting why each slice lives where it does — otherwise six months later every new feature defaults to the noisiest path and you lose the governance or latency benefits you paid for.

If you are consolidating vendors for procurement simplicity, remember that “one invoice” on AWS can still mean multiple Bedrock SKUs plus transfer. Conversely, “one Anthropic invoice” can still require a separate cloud for the rest of your stack. Optimize for operational clarity and net dollars after discounts, not logo count.

Practical checklist before you commit

1. Confirm the exact model ID and whether you will call global vs regional endpoints.
2. Ask finance whether an AWS EDP or Anthropic commercial deal applies to Claude spend.
3. Inventory add-ons: Guardrails, Knowledge Bases, AgentCore, provisioned throughput, and expected cross-region transfer.
4. Separate traffic that needs Fast mode or first-party betas from traffic that can stay on Bedrock.
5. Prototype Batch + caching on the path you choose; measure cache_read_input_tokens and async completion times before locking a forecast. Use the cost calculator for rough volume math, then meter reality.
6. Do not confuse Claude Platform on AWS (CCUs) with classic Bedrock on-demand when reading Marketplace invoices. [VERIFY: 2026-09-19]

Token hygiene still dominates either path. Cutting prompt bloat, tool-loop chatter, and unnecessary long outputs often saves more than switching clouds — start with cut token usage and how tokens are billed.

One more forecasting tip: build scenarios at global list, regional +10%, and Batch 50% off before you layer EDPs or cache hit rates. That three-row table usually surfaces whether residency or async eligibility matters more than the Bedrock vs API brand choice. Teams that skip the table often over-index on list parity and under-index on the levers that actually move the invoice.

Get a written Bedrock vs API cost review

If you are choosing between Amazon Bedrock and the Anthropic Claude API — or running both and want a cleaner split — paste your model mix, region pins, Batch share, and AWS discount context into a Claude token audit. You get a written review of where list parity holds, where the ~10% residency premium and add-on SKUs bite, and which workload slices should stay first-party for Fast mode or new features. Soft CTA only on this page — no audit dollar price here.