Migrate from Anthropic’s Claude 3.5 Sonnet to Claude 4 Sonnet on Amazon Bedrock

Anthropic’s Claude 4 Sonnet is designed to comply with directions extra exactly. It is a main benefit, however would possibly imply it’s much less verbose or elaborative than Claude 3.5 Sonnet until explicitly prompted to be. Your system prompts defining the AI’s persona would possibly want changes.

New reasoning options

Anthropic’s Claude 4 Sonnet mannequin’s built-in prolonged pondering is a strong functionality. To unlock the deep reasoning capabilities of the brand new fashions, builders can allow prolonged pondering by together with the pondering key phrase argument of their API name. However be certain to make use of it strategically—prolonged pondering incurs extra prices as a result of reasoning tokens are billed as output tokens at commonplace mannequin charges. You’ll be charged for the complete pondering course of (not simply the abstract you see), and prices can scale with the pondering finances you set.

We advocate enabling prolonged pondering for duties requiring deep, multi-step evaluation the place accuracy is paramount (for instance, complicated scientific issues or superior coding challenges).

We advocate disabling prolonged pondering for easy queries or latency-sensitive functions. The characteristic provides to each value (reasoning tokens are billed) and response time. For a lot of duties, a well-crafted CoT immediate remains to be probably the most environment friendly alternative.

To make use of prolonged pondering, merely replace the Converse API name by including the additionalModelRequestFields parameter with the pondering configuration. You possibly can specify the utmost pondering tokens that the mannequin can use by setting a worth for budget_tokens. Notice that maxTokens needs to be bigger than budget_tokens for prolonged pondering. See the next code:

import boto3

bedrock_runtime = boto3.consumer(service_name="bedrock-runtime")

# Add configurations to allow prolonged pondering
response = bedrock_runtime.converse(
    modelId='us.anthropic.claude-sonnet-4-20250514-v1:0',
    messages=[{'role': 'user', 'content': [{'text': "Your prompt here"}]}],
    inferenceConfig={'maxTokens': 2048},
    additionalModelRequestFields={"pondering":{"kind": "enabled",
     "budget_tokens": 1024}}
)

reasoning_text = response['output']['message']['content'][0]['reasoningContent']['reasoningText']['text']
assistant_response = response['output']['message']['content'][1]['text']

The next desk exhibits an instance of prolonged pondering and the way it can enhance response accuracy. We immediate Anthropic’s Claude 3.5 Sonnet v2 and Claude 4 Sonnet (with pondering enabled) with an instance immediate:

Alan, Bob, Colin, Dave and Emily are standing in a circle. Alan is on Bob's instant left. Bob is on Colin's instant left. Colin is on Dave's instant left. Dave is on Emily's instant left. Who's on Alan's instant proper?

The desk exhibits responses from the 2 fashions. With prolonged pondering, the mannequin is ready to fastidiously purpose by the issue and arrive on the appropriate reply.

With prolonged pondering enabled, the API will, by default, return a abstract of the mannequin’s full thought course of slightly than the complete chain of reasoning. It is a safeguard designed to stop misuse whereas nonetheless offering the complete intelligence advantages. Builders needs to be conscious that this can lead to a chunky or delayed sample in streaming responses and may account for this of their software’s person expertise.

Anthropic’s Claude 4 fashions characteristic the interleaved pondering functionality when utilizing prolonged pondering with instruments, permitting the mannequin to carry out intermediate reasoning between software calls and develop extra nuanced analytical conclusions primarily based on the software outcomes obtained. To allow interleaved pondering for software calling, add the extra parameter "anthropic_beta": ["interleaved-thinking-2025-05-14"] to additionalModelRequestFields within the Converse API request.

For extra particulars on utilizing prolonged pondering, consult with Prolonged pondering.

Strong analysis is non-negotiable

You will need to confirm that the brand new mannequin performs in addition to or higher than the outdated one in your particular duties. Create a curated set of prompts and anticipated outputs which can be extremely consultant of your manufacturing site visitors. This tradition benchmark is way extra priceless than reported benchmark scores.

Combine this dataset into an automatic analysis pipeline. This pipeline ought to change into a part of your steady integration and steady supply (CI/CD) course of, serving as a everlasting regression suite for future mannequin or immediate modifications. For tooling, groups can use built-in options like Amazon Bedrock evaluations or open supply analysis frameworks like RAGAS and DeepEval.

Managing built-in security and governance

A mannequin’s security profile modifications with every model. This should be examined as a part of an built-in system. By no means take a look at the brand new mannequin in isolation. Your migration take a look at plan is simply legitimate if it evaluates the brand new mannequin with the very same guardrail configuration you utilize in manufacturing, for instance, Amazon Bedrock Guardrails. A brand new mannequin’s conversational type would possibly set off current guardrails in surprising methods, resulting in a spike in blocked responses.

Implementing protected deployment methods

When deploying Anthropic’s Claude 4 Sonnet to manufacturing, implement a phased rollout technique to reduce dangers. Contemplate shadow testing to match mannequin efficiency utilizing mirrored site visitors with out impacting customers, then conduct A/B testing to measure impression on enterprise KPIs. For the precise rollout, use both a canary launch strategy by progressively exposing a small proportion of customers to the brand new mannequin, or a blue/inexperienced deployment technique that maintains parallel environments for fast rollback functionality. This structured strategy facilitates protected validation of the brand new mannequin’s efficiency whereas sustaining enterprise continuity.

Conclusion

By treating the transition from Anthropic’s Claude 3.5 Sonnet (v1 and v2) to Claude 4 Sonnet as a structured engineering mission, you possibly can de-risk the migration course of and unlock vital advantages. Understanding the important thing mannequin variations, adapting your prompts and API calls, and implementing a strong, automated analysis technique are the pillars of a profitable improve.

This transition is a chance to take care of enterprise continuity and improve your software with next-generation capabilities. We encourage you to start your evaluation and testing as we speak.

For extra particulars, consult with Migrating to Claude 4 and Anthropic’s Claude in Amazon Bedrock, and attain out to us for migration-specific inquiries.

Concerning the authors

Melanie Li, PhD, is a Senior Generative AI Specialist Options Architect at AWS primarily based in Sydney, Australia, the place her focus is on working with prospects to construct options utilizing state-of-the-art AI/ML instruments. She has been actively concerned in a number of generative AI initiatives throughout APJ, harnessing the facility of LLMs. Previous to becoming a member of AWS, Dr. Li held knowledge science roles within the monetary and retail industries.

Deepak Dalakoti, PhD, is a Deep Studying Architect on the Generative AI Innovation Centre in Sydney, Australia. With experience in AI, he companions with shoppers to speed up their generative AI adoption by personalized, progressive options. Exterior the world of AI, he enjoys exploring new actions and experiences.

Mahsa Paknezhad, PhD, is a Deep Studying Architect on the AWS Generative AI Innovation Heart. With a deal with scalability and manufacturing readiness, Mahsa helps organizations throughout various industries harness superior generative AI fashions to attain significant enterprise outcomes. Mahsa has a confirmed monitor file of efficiently delivering tasks particularly inside the mining business and the healthcare sector.

Nicholas Moore is a Options Architect at AWS, serving to startups and companies of all sizes flip concepts into actuality. He focuses on cloud options with a deal with AI, analytics, and trendy software improvement.

Derrick Choo is a Senior Options Architect at AWS who accelerates enterprise digital transformation by cloud adoption, AI/ML, and generative AI options. He focuses on full-stack improvement and ML, designing end-to-end options spanning frontend interfaces, IoT functions, knowledge integrations, and ML fashions, with a specific deal with pc imaginative and prescient and multi-modal techniques.

Sovik Kumar Nath is an AI/ML and Generative AI Senior Options Architect with AWS. He has in depth expertise designing end-to-end ML and enterprise analytics options in finance, operations, advertising, healthcare, provide chain administration, and IoT. He has grasp’s levels from the College of South Florida and College of Fribourg, Switzerland, and a bachelor’s diploma from the Indian Institute of Know-how, Kharagpur. Exterior of labor, Sovik enjoys touring, taking ferry rides, and watching films.

Saurabh Trikande is a Senior Product Supervisor for Amazon Bedrock and Amazon SageMaker Inference. He’s obsessed with working with prospects and companions, motivated by the objective of democratizing AI. He focuses on core challenges associated to deploying complicated AI functions, inference with multi-tenant fashions, value optimizations, and making the deployment of generative AI fashions extra accessible. In his spare time, Saurabh enjoys mountaineering, studying about progressive applied sciences, following TechCrunch, and spending time together with his household.

Gareth Jones is a Product Supervisor at Anthropic targeted on the Claude API. His work contains collaboration with AWS on Amazon Bedrock to make Claude broadly accessible to builders.