The final countdown for Claude 3.5 Sonnet

I was reading the latest AWS release notes when one line made me stop and reread.

That caught my attention because many of my AI workflows, such as multi-step agents, automated summaries, and background jobs, still run on Claude 3.5 Sonnet through Bedrock. My first instinct was, “No rush, I’ll handle it later.” But the more I thought about it, the more the risk became obvious. This was not a minor version bump, and ignoring it could break everything.

If you run AI workloads on Amazon Bedrock with Claude 3.5 Sonnet, its retirement will impact your workflows as much as mine. So, unless you start planning the migration now, you will face the same risks I did.

This newsletter will discuss the key changes developers need to know, the migration challenges you will likely face, and the steps you can take to move without disrupting your workflows. By the end, you will know how to plan, test, and deploy Claude Sonnet 4 so your AI systems keep running smoothly and take advantage of its new capabilities.

How the retirement impacts your Bedrock workloads#

Even though Claude 3.5 Sonnet is still answering API calls, its “legacy” status on Bedrock already changes a few things behind-the-scenes. It no longer receives performance or reliability updates, and its support window is shrinking. That means if you’re running multi-step agents, batch summarization jobs, chatbots, or any other automated flow on Bedrock, you’re already operating on a model that won’t be fixed if something breaks.

When the endpoint finally shuts off in March 2026, Bedrock will not automatically route your calls to Claude Sonnet 4. Unless you migrate, those agents and jobs will start throwing errors. The sooner you test Claude Sonnet 4 against your current workloads, the easier it will be to smooth out responses, latency, and cost differences before the deadline.

Migrating from Claude 3.5 Sonnet to Claude Sonnet 4 on Bedrock#

Now that you know why retirement matters, the next question is what to do about it. Migrating early gives you control over the process instead of scrambling close to the deadline. Below is a refined playbook that gives you a systematic way to migrate.

Step 1: Take inventory of every 3.5 integration#

Search your codebase, infrastructure-as-code (IaC) templates, environment configurations, and the Bedrock console for any references to anthropic.claude-3-5-sonnet (v1 or v2). Ensure you include less obvious places such as cron jobs, background workers, and batch pipelines.

Doing this will give you a comprehensive list of everything that will stop working once 3.5 is retired.

Step 2: Document each workload’s assumptions#

For each integration you identify, document exactly how it depends on Claude 3.5 Sonnet. Note the prompt structure it expects, the output format your parsing logic relies on, token limits, latency profile, and cost assumptions. This documentation becomes the test plan you will use when validating Claude Sonnet 4, so you are not surprised by differences.

Step 3: Create a parallel test environment#

Now, you can safely set up a separate environment to test your workloads against Claude Sonnet 4 without disrupting production. Set up a second Bedrock integration that points to anthropic.claude-sonnet-4-20250514-v1:0. Replay real requests into it or route a small slice of live traffic so you can measure differences in responses, speed, and cost without impacting production. Running both models side by side allows you to catch breaking changes early.

Note: If you’d like a hands-on example of setting up Bedrock integrations in a serverless environment, check out our Building Serverless GenAI Applications with Amazon Bedrock Cloud Lab. It walks you through creating a complete serverless AI application, combining Bedrock models, Guardrails, Lambda, API Gateway, and Aurora to deliver end-to-end intelligent query processing.

Step 4: Adapt and validate#

Based on the test results, update your prompts, post-processing logic, and timeouts where necessary. Claude Sonnet 4 introduces changes such as new stop reasonsIn Claude models, a stop reason indicates why the model ended its response. For example, reaching a token limit, completing the task naturally, or stopping due to safety or policy rules. and removing the undo command, so your error-handling code must be updated to cope with these differences. Run your automated tests and compare outputs until satisfied with the results.

Step 5: Plan a staged rollout#

Decide how you will move production traffic to Claude Sonnet 4. You might shift the workload incrementally, by task, region, or by increasing traffic volume. Update your CI/CD pipelines with the new model ID, but keep 3.5 running alongside 4 for a short overlap, so you can roll it back instantly if an issue appears.

Step 6: Monitor and retire 3.5#

Add alerts and dashboards to track the new model’s behavior, including error rates, refusals, latency anomalies, and divergence in output quality. Once you are confident everything runs smoothly, remove all references to 3.5 well before March 2026 so that nothing fails when the endpoint disappears.

Unlocking new capabilities with Claude Sonnet 4#

Claude Sonnet 4 brings improvements that can transform your workflows. The context window has grown from 200,000 tokens to 1 million (beta), letting you handle entire codebases, long financial reports, or legal documents in a single prompt. This simplifies complex multi-step workflows and reduces the need to chunk data artificially.

Claude Sonnet 4 also introduces built-in reasoning features. Unlike 3.5, which relied on chain-of-thought prompts, you can now enable extended and interleaved thinkingInterleaved thinking in Claude 4 lets the model pause mid-response to reason, execute tools, and integrate their results before continuing its output.. These give the model dedicated computation time to reason before answering, improving performance on complex problems. You can selectively enable extended thinking for tasks that require deep analysis, while keeping simple queries fast and cost-efficient. Extended thinking can be enabled through the API with the additionalModelRequestFields parameter. This parameter should be structured as follows:

Note: Keep in mind that reasoning tokens are billed, so use these features strategically. Enable them for complex, multi-step tasks and disable them for simple, latency-sensitive queries.

Claude Sonnet 4’s tool usage has been significantly upgraded. Unlike Claude 3.5 Sonnet, which executes tools one at a time in sequence, Claude Sonnet 4 can run multiple tools simultaneously and interleave reasoning between calls. For example, an agentic workflow that gathers financial data, performs a risk analysis, and then generates a report can now execute these steps in parallel, with the model reasoning over partial outputs between tool calls. This reduces total execution time and improves accuracy because the model can integrate intermediate results into its final output. By leveraging parallel and interleaved tool use, developers can build faster, reliable, and higher-quality AI workflows without manually orchestrating each step.

Final thoughts and looking ahead#

Migrating from Claude 3.5 Sonnet to Claude Sonnet 4 is an opportunity to strengthen your AI workflows and avoid last-minute crises. By taking inventory of all 3.5 integrations, understanding each workload’s assumptions, testing in parallel, and planning a staged rollout, you can migrate smoothly without compromising on the production.

Looking ahead, keep in mind that once a model launches on Amazon Bedrock, it will remain available for at least 12 months before its official end-of-life. Model updates and new releases are frequent, so staying informed is crucial. Regularly check the Bedrock model life cycle documentation to track EOL dates and plan future migrations proactively. This habit ensures that your AI workflows remain reliable and resilient as the platform evolves.

To further enrich your learning, explore the following Cloud Labs:

• https://www.educative.io/cloudlabs/building-rag-chatbot-using-langchain-and-amazon-bedrock

• https://www.educative.io/cloudlabs/deploying-a-machine-learning-model-with-amazon-sagemaker