Claude Haiku 5.5 on Amazon Bedrock

Quick Answer

Yes. AWS announced Claude Haiku 5.5 availability on October 7, 2026. Bedrock and Claude Platform on AWS are distinct access paths; this guide uses Bedrock Runtime with boto3 Converse.

Start with an authenticated AWS role, a supported source region and a compatible inference profile. Review routing and data residency before sending application data. No API call was made by this site.

Model IDs and Inference Profiles

The foundation model identifier names the model; a profile supplies routing. Do not add a guessed version suffix. Bedrock Runtime’s regional availability table lists cross-region inference for this model, not an in-region runtime endpoint.

Supported Regions

Selected commercial source regions are shown below, not a complete region list. All three support geo and global Runtime routing; in-region Runtime inference is not listed. Verified October 9, 2026. Check the live model card before changing regions.

Python boto3 Example

Install boto3 in your application environment. Authenticate using an AWS role or an existing AWS IAM Identity Center profile and the standard credential chain. Never embed access keys in source control. This example uses a US geo profile with us-east-1; change both settings together only after checking AWS support.

Converse accepts messages containing content blocks and returns output.message.content, usage and stopReason. A token limit can truncate a response; inspect stopReason before treating an answer as complete. This non-streaming example is for request wiring, not latency measurement.

import os
import boto3
from botocore.exceptions import BotoCoreError, ClientError

# Uses boto3's standard credential chain; no keys in this file.
region = os.getenv("AWS_REGION", "us-east-1")
model_id = os.getenv(
    "BEDROCK_INFERENCE_PROFILE_ID", "us.anthropic.claude-haiku-5-5"
)
client = boto3.client("bedrock-runtime", region_name=region)

def main():
    try:
        response = client.converse(
            modelId=model_id,
            messages=[{
                "role": "user",
                "content": [{"text": "Classify: My shipment has not arrived. Reply with a short label."}],
            }],
            inferenceConfig={"maxTokens": 256},
        )
        blocks = response.get("output", {}).get("message", {}).get("content", [])
        print("".join(block.get("text", "") for block in blocks))
        print("Token usage:", response.get("usage", {}))
        print("Stop reason:", response.get("stopReason", "unknown"))
    except ClientError as exc:
        error = exc.response.get("Error", {})
        code = error.get("Code", "Unknown")
        if code == "AccessDeniedException":
            hint = "Check model access, IAM, profile resources and destination-region policies."
        elif code == "ValidationException":
            hint = "Check request fields, source region and inference profile ID."
        else:
            hint = "Consult the Converse error reference; do not blindly retry."
        request_id = exc.response.get("ResponseMetadata", {}).get("RequestId", "unknown")
        raise SystemExit(f"{code}: {hint} Request ID: {request_id}") from exc
    except BotoCoreError as exc:
        raise SystemExit("Check AWS credentials, connection and SDK configuration.") from exc

if __name__ == "__main__":
    main()

Request/response structure follows AWS Converse. Offline syntax and mocked response/error checks only; not tested against the live API. Running it with AWS credentials can incur charges.

Pricing and Billing

AWS’s model card identifies Marketplace billing: charges appear under the provider in the AWS bill and Cost Explorer. The card lists Standard support; do not assume general Bedrock Batch, Flex or Priority offerings apply to this model.

Haiku 5.5 AWS input/output unit rates: Not publicly confirmed in the accessible pricing content reviewed October 9, 2026. Verify the selected region, routing, context tier, cache conditions, currency and contract in AWS’s pricing view before budgeting. Anthropic’s direct API prices are not a verified AWS quote.

Once AWS rates are verified in USD per million tokens, estimate token cost as inputTokens / 1,000,000 × AWS inputRate + outputTokens / 1,000,000 × AWS outputRate. Calculate applicable tiers separately; add cache, tools, retries, ancillary services and taxes as applicable. This formula is arithmetic, not a measured invoice.

Troubleshooting

Diagnose the returned error code and request ID before retrying. An unsupported source region or profile does not imply that the model is unavailable everywhere. Do not broaden permissions to all resources simply to suppress an error.

Related Developer Guides

Frequently asked questions