What to Do Before AWS Bedrock Shuts Down Legacy Claude 3.x Models on November 13, 2026

Shutdown: Nov 13, 2026 (Thursday). AWS Bedrock ends five legacy Anthropic models in the listed Regions: Claude 3 Sonnet, Claude 3.5 Sonnet (v1 and v2), Claude 3.7 Sonnet and Claude 3 Haiku. In ca-central-1, Claude 3 Sonnet and Claude 3 Haiku follow on Nov 27, 2026. After EOL, requests fail. 2026-11-13

Affected model IDs (from AWS)

• Source: AWS Bedrock Model lifecycle (Legacy), "Models with a scheduled EOL date", read Nov 10, 2026. Bedrock dates can differ from Anthropic's own API dates; for Bedrock, only AWS's dates apply.
• Regions matter: AWS lists these dates per Region. A Region not in the row is not covered by this announcement, and that does not mean the model stays available there. Check your Region (Step 2).
• Before EOL: AWS warns existing customers may lose access to Legacy models after 15 days of inactivity, and you can't create new Provisioned Throughput on a Legacy model.
• Track every Bedrock date on the AWS Bedrock page or the shutdown calendar.

Recommended replacements

AWS's lifecycle page does not name a replacement. Anthropic's model deprecations page maps the same models (on its own API) like this, and both targets are Active on Bedrock:

Newer option for Haiku workloads: Claude Haiku 5.5 (anthropic.claude-haiku-5-5, profiles us./ca./eu./au./jp./global.; card). Anthropic's mapping names Haiku 4.5, so test both if cost matters.

Step 1: Find every place you call these IDs

rg -n --hidden -g '!node_modules' -e 'claude-3-(5-|7-)?sonnet' -e 'claude-3-haiku' .
rg -n --hidden -g '*.{yml,yaml,json,toml,tf,tfvars}' -g '.env*' -e 'anthropic\.claude-3' .

• Catch prefixed forms too: us.anthropic.claude-3-5-sonnet-20241022-v2:0, apac.anthropic..., eu.anthropic... and inference-profile ARNs.
• Check Lambda env vars, Step Functions, Bedrock Agents and Knowledge Base configs, Flows, Prompt management, CloudFormation/CDK/Terraform and LLM gateway routing tables.

Step 2: Confirm the lifecycle in each Region you use

for r in ap-south-1 ap-southeast-1 ap-northeast-2 eu-west-2 eu-central-2 sa-east-1 ca-central-1; do
  echo "$r: $(aws bedrock get-foundation-model --region $r \
    --model-identifier anthropic.claude-3-5-sonnet-20241022-v2:0 \
    --query 'modelDetails.modelLifecycle.status' --output text 2>&1)"
done
aws bedrock list-inference-profiles --region ap-south-1 \
  --query "inferenceProfileSummaries[?contains(inferenceProfileId,'sonnet-5-5')].inferenceProfileId"

Swap in each model ID you found in Step 1. AWS: "To check the current state for the Region you use, call ListFoundationModels or GetFoundationModel in that Region and read the modelLifecycle field."

Step 3: Swap the model ID (boto3)

Keep the ID in config so rollback is one change. Converse API (same code for any Claude model):

import os, boto3
brt = boto3.client("bedrock-runtime", region_name=os.environ.get("AWS_REGION", "ap-south-1"))
# was: "anthropic.claude-3-5-sonnet-20241022-v2:0"
MODEL_ID = os.environ.get("BEDROCK_MODEL_ID", "global.anthropic.claude-sonnet-5-5")

resp = brt.converse(
    modelId=MODEL_ID,
    messages=[{"role": "user", "content": [{"text": "Summarize this ticket in 3 bullets: ..."}]}],
    inferenceConfig={"maxTokens": 1024},
)
# Newer Claude models can return reasoning blocks first: read only the text blocks.
text = "".join(b["text"] for b in resp["output"]["message"]["content"] if "text" in b)
print(text, resp["usage"])

InvokeModel (Anthropic Messages body):

import json
body = {"anthropic_version": "bedrock-2023-05-31", "max_tokens": 1024,
        "messages": [{"role": "user", "content": "Summarize this ticket in 3 bullets: ..."}]}
resp = brt.invoke_model(modelId=MODEL_ID, body=json.dumps(body))
out = json.loads(resp["body"].read())
text = "".join(b.get("text", "") for b in out["content"] if b["type"] == "text")

Step 4: Region and inference-profile notes

• Haiku 4.5 needs a profile ID: AWS says the bare model ID isn't supported for on-demand throughput on bedrock-runtime. Use a geo (us., eu., au., jp., in.) or global. profile ID. From ap-south-1 (Mumbai), the in. profile is listed on the card.
• Sonnet 5.5 geo profiles are listed for US, EU and AU only. From Mumbai, Singapore, Seoul or São Paulo, check the card's Regional availability table for your source Region; global. is the listed cross-Region option there.
• Data residency: geo and global profiles can route requests to other Regions. If you need single-Region processing, confirm In-Region support for your Region on the model card before switching.
• IAM: calling a profile needs bedrock:InvokeModel (or InvokeModelWithResponseStream) on the inference-profile ARN and on the foundation-model ARNs in the destination Regions. Update policies that were scoped to the old model ARN.
• Quotas: on-demand quotas are per model and profile. Check Service Quotas for the new model before moving production traffic.
• Provisioned Throughput / custom models: AWS says you can't create new Provisioned Throughput for Legacy models; plan any PT purchase on the new model.

Step 5: Testing checklist

• Run a frozen set of real prompts through old and new models; compare quality, refusals, length and latency.
• Thinking is on by default on Claude Sonnet 5.5 (adaptive thinking, default effort high; the card says thinking.type: "disabled" is not supported). Re-check latency, output-token cost and any code that assumes the first content block is text.
• Tool use: validate every tool call against your schemas and re-run multi-step agent loops end to end.
• Structured output: parse 100% of responses with your production JSON parser.
• Limits: re-check maxTokens, stop sequences, temperature settings and context-size assumptions against the new model card.
• Cost: compare per-token prices on the Bedrock pricing page using your own token counts.
• Errors: alert on ValidationException / ResourceNotFoundException and access-denied errors, which is typically how a retired ID or a missing IAM grant shows up.

Step 6: Roll out before Nov 13 and block the old IDs

• Canary 5–10% of traffic first, then 100%, with the old ID one config change away while it still works.
• After cutover, add a CI check that fails if any of the five anthropic.claude-3* IDs above appears in the repo.
• Using Claude Sonnet 4.5 on Bedrock too? Its Bedrock EOL is Apr 8, 2027 in most Regions, not Anthropic's Nov 30 API date (Sonnet 4.5 guide).

FAQ

When do the legacy Claude 3.x models shut down on AWS Bedrock?

AWS lists an EOL date of November 13, 2026 for anthropic.claude-3-5-sonnet-20240620-v1:0, anthropic.claude-3-5-sonnet-20241022-v2:0, anthropic.claude-3-7-sonnet-20250219-v1:0, anthropic.claude-3-sonnet-20240229-v1:0 and anthropic.claude-3-haiku-20240307-v1:0 in the regions on its Model lifecycle (Legacy) page. In ca-central-1, Claude 3 Sonnet and Claude 3 Haiku end on November 27, 2026.

Which regions are affected?

Per AWS: ap-northeast-2, ap-south-1, ap-southeast-1, eu-west-2, eu-central-2 and sa-east-1, depending on the model (see the table), plus ca-central-1 on Nov 27 for Claude 3 Sonnet and Claude 3 Haiku. AWS notes that the Regions column lists only Regions covered by the announced date, so check modelLifecycle in your own Region.

What should I replace them with?

AWS does not name a replacement on its lifecycle page. Anthropic's own deprecations page maps Claude 3 Sonnet, 3.5 Sonnet and 3.7 Sonnet to Claude Sonnet 5.5 and Claude 3 Haiku to Claude Haiku 4.5. Both are Active on Bedrock.

What happens after the EOL date?

AWS says the model is no longer available in the listed Regions and requests to it fail. Migration does not happen automatically; you have to change the model ID in your code.

Can I keep using a Legacy model until the last day?

Yes, but AWS warns that existing customers may lose access to Legacy models after 15 days of inactivity, and you cannot create new Provisioned Throughput for Legacy models.

Dates come from official provider pages and are re-checked daily. Always confirm on the linked source. Related: all AWS Bedrock shutdowns · Anthropic API shutdowns · shutdown calendar · all guides