Claude on Amazon Bedrock: models, prices and how to start.

IWhich Claude models are on Amazon Bedrock?

Claude comes in three sizes: Haiku (fast and cheap), Sonnet (the default for most work) and Opus (the most capable). Bedrock also carries two specialist lines and older versions.

On October 10, 2026, the AWS list of Anthropic models showed 19 Claude models. Start with the three 5.5 models. The details below come from each AWS model card.

IIHow much does Claude cost on Bedrock?

Standard rates per 1 million tokens, as listed for US East (Ohio) on October 10, 2026. Prices change often.

Source: Amazon Bedrock pricing. Global cross-Region inference can send a request to any Region. Geo and In-Region keep it in one geography or Region and cost 10% more on these models. Caching and the other ways to pay are on our Bedrock pricing page.

Guides that quote $0.25 and $1.25 for Haiku, or $15.00 and $75.00 for Opus, are quoting Claude 3 Haiku and Claude Opus 4 and 4.1.

IIIWhich Claude model should you use?

Start with the cheapest model that passes your own test, and move up only the requests that fail.

• Claude Haiku 5.5: High-volume work where speed matters: classification, extraction, routing, short answers. AWS says it beats Haiku 4.5 at coding and agent tasks.
• Claude Sonnet 5.5: The default for most product features: support agents, document questions, coding help and writing.
• Claude Opus 5.5: Hard reasoning, long-running agents and large code changes. It costs twice Sonnet 5.5 per token.

For scale: 1,000 requests with 3,000 input and 800 output tokens each cost about $0.70 on Haiku 5.5, $14.00 on Sonnet 5.5 and $28.00 on Opus 5.5 at Global list prices, before caching and thinking. Moving a request between models is a change of model ID.

IVHow do you get access to Claude on Bedrock and make the first call?

There is no Model access page to click through in commercial Regions any more. Access turns on with your first call, if three things are in place.

• The IAM role that makes the first call has the AWS Marketplace permissions aws-marketplace:Subscribe, Unsubscribe and ViewSubscriptions. Bedrock then subscribes the account in the background, which can take up to 15 minutes.
• The account has a valid payment method for AWS Marketplace purchases.
• You have sent Anthropic's one-time use-case form, once per account or once at the AWS organization's management account, from the model catalog or with the PutUseCaseForModelAccess API.

Then call a model ID. global.anthropic.claude-sonnet-5-5 uses the Global endpoint. Geo IDs such as us.anthropic.claude-sonnet-5-5 keep requests in one geography. The Converse API, the Invoke API and the Anthropic Messages format all work. A minimal Converse call in Python:

import boto3
client = boto3.client("bedrock-runtime", region_name="us-east-1")
resp = client.converse(
  modelId="global.anthropic.claude-haiku-5-5",
  messages=[{"role": "user", "content": [{"text": "Summarize this clause: ..."}]}],
  inferenceConfig={"maxTokens": 512},
)
print(resp["output"]["message"]["content"][0]["text"])

Getting AccessDeniedException after sending the form? The agreement may still be pending. Check it with GetFoundationModelAvailability (AWS model access docs).

VShould you use Bedrock or the Anthropic API?

It is the same Claude either way. The difference is where it runs, how you pay and when features arrive.

• Security: Bedrock uses IAM roles, AWS PrivateLink and CloudTrail. There is no separate Anthropic API key to store and rotate.
• Data: AWS says Bedrock content is not used to improve the base models and is not shared with model providers (Bedrock FAQs). Use a Geo model ID if data must stay in one geography.
• Billing: Claude charges land on your AWS bill, and AWS Activate credits can pay them. Anthropic's Claude for Startups credits cannot.
• Features: On October 10, 2026, bedrock-runtime did not support structured outputs or token counting for the Claude 5.5 models, and they ran on the Standard tier only. New Anthropic features can reach the Anthropic API first. Knowledge Bases, Agents, Flows and Guardrails do work with them.

Pick the Anthropic API if you are not on AWS, need a feature Bedrock lacks, or hold Anthropic credits. Pick Bedrock if your stack runs on AWS and you want one bill, IAM, and AWS credits or partner funding to cover the spend.

VIHow does Claude compare with Nova, GPT, Llama and Mistral on Bedrock?

Bedrock offers many model families through one API, so you can test them side by side on your own prompts.

Amazon Nova costs less per token. Per million tokens, Nova Micro is $0.035 in and $0.14 out; Nova Pro is $0.80 in and $3.20 out. Many teams use Nova for simple, high-volume steps and Claude where answer quality matters (Nova vs Claude).

OpenAI models are on Bedrock too, from the open-weight gpt-oss-120b ($0.15 in and $0.60 out) to newer GPT models priced on their model cards. Llama 4 Maverick ($0.24 in and $0.97 out) and Mistral Large 3 ($0.50 in and $1.50 out) are open-weight options. Google's Gemini models are not on Bedrock; its open Gemma models are.

To choose, run 50 to 100 of your real prompts through two or three models and compare cost per correct answer, not cost per token.

VIIWho pays for Claude on Bedrock if you are a startup?

Three sources of money can cover it, and each works differently.

Claude for Startups credits from Anthropic work only on Anthropic's own API, not on Bedrock (see our Claude for Startups page).

AWS Activate credits can pay for Claude on Bedrock (AWS announcement, April 2024; Activate terms). You apply yourself: AWS Activate Founders gives $1,000 to start, up to $5,000, and AWS Activate Portfolio goes up to $200,000 with your investor's or accelerator's Org ID. A consulting partner cannot apply for you.

Partner-filed AWS funding is the larger route for most startups without a VC. A vetted AWS partner can file POC funding to test a Claude workload, IW funding when you already run on AWS and launch an AI feature, or MAP funding if you move to AWS. For startups it is usually $50K to $150K. Depends on the workload; AWS approves.

VIIIWhat can you do without a partner?

A lot. Many teams never need one.

• Apply for AWS Activate Founders at aws.amazon.com/startups/credits. AWS says Activate applicants hear back within 5 to 10 business days.
• Cache long system prompts and tool definitions.
• Use the Global model ID unless data must stay in one geography. It is 10% cheaper.
• Send easy requests to Haiku 5.5 or Nova, and keep Opus for the hard ones.
• Estimate with the AWS Pricing Calculator and set an AWS Budgets alert.

You do not need a partner if Activate covers your bill and you are not testing a larger workload, launching new work on AWS or moving to AWS.