AWS Generative AI Developer Professional (AIP-C01) guide

AWS Certified Generative AI Developer – Professional (exam AIP-C01) tests whether you can build production applications on top of existing foundation models: retrieval-augmented generation, agents, guardrails, evaluation and cost control, mostly on Amazon Bedrock. It is a professional exam, with 75 questions in 180 minutes for 300 USD, and it has been out of beta since March 2026.

This guide explains what the exam covers as of October 2026: its format and history, each of the five domains in plain terms, the services you need to know (and the ones that are named but closed to new customers), how it compares with AI Practitioner and Machine Learning Engineer, and how to prepare. Every exam fact was checked on AWS's certification page and exam guide on October 21, 2026.

AIP-C01 at a glance

As of October 2026:

Full name: AWS Certified Generative AI Developer – Professional, exam code AIP-C01.
Questions: 75, of which 65 are scored and 10 are unscored. You can't tell which are which.
Question types: multiple choice and multiple response. A multiple-response question earns credit only if you pick every correct option.
Time: 180 minutes.
Passing score: 750 on a scale of 100 to 1,000. Scoring is compensatory, so you don't need to pass each domain separately, and there is no penalty for guessing.
Price: 300 USD, plus any foreign-exchange charges.
Delivery: at a Pearson VUE test centre or online with a remote proctor.
Languages: English, Japanese, Korean and Simplified Chinese.
Validity: three years, like every AWS certification.

There is no required prerequisite. AWS recommends two or more years of building production applications on AWS or with open-source technologies, plus a year of hands-on generative AI work. It also expects some background in AI/ML or data engineering, and working knowledge of compute, storage, networking, security, infrastructure as code, monitoring and cost.

The official details are on the Generative AI Developer – Professional certification page and in the AIP-C01 exam guide.

From beta to the standard exam

AWS announced the exam on October 14, 2025 and ran it as a beta first. The last day to take the beta was March 31, 2026. In an update dated March 17, 2026, AWS said registration was open for the standard version and that the exam had been refreshed to add Amazon Bedrock AgentCore, its service for running agents in production.

That history matters when you pick study material. Some prep sites still describe the beta, so go by the 75 questions and 180 minutes AWS lists today. Material written before the refresh may also say little about AgentCore. Check that a course or practice set covers the standard AIP-C01 exam and the current exam guide.

The five domains in practice

The exam guide splits the content into five domains. Here is what each one means for a developer, with practice questions from the AWS AI Practitioner bank on the same services. Those questions are written for the foundational exam, so they are a starting point: the professional exam asks about the same features in longer scenarios with more constraints.

Foundation model integration, data management and compliance (31%)

The largest domain is about everything around the model before a request reaches it. It covers analysing requirements and designing a solution (including a proof of concept on Bedrock and the Well-Architected Generative AI Lens), selecting and configuring models (switching models without redeploying, cross-Region inference, deploying a LoRA fine-tuned model), and building data pipelines that validate and format input for a model.

Two of the domain's six tasks are about retrieval. You need to design vector stores with Bedrock Knowledge Bases, OpenSearch, Aurora with pgvector, and DynamoDB for metadata, then tune retrieval with chunking, embeddings, hybrid search and rerankers. A good first check is this practice question on where a knowledge base stores its embeddings. The domain closes with prompt engineering and governance: Bedrock Prompt Management, Prompt Flows, Guardrails and regression tests for prompts.

If retrieval is new to you, read what RAG is and how it works before the exam guide.

Implementation and integration (26%)

This domain is about wiring models into real systems. It starts with agentic AI: building agents with frameworks such as Strands Agents and AWS Agent Squad, exposing tools through MCP servers on Lambda or ECS, and adding human review where an agent's action is risky. Then deployment: on-demand calls through Lambda, Bedrock provisioned throughput, SageMaker AI endpoints, and cascading from a small model to a larger one.

The rest is integration work: API Gateway and EventBridge in front of models, synchronous and asynchronous calls, streaming, back-off and rate limits, routing between models, CI/CD for generative AI apps, and tracing with X-Ray. This practice question on hosting agents in production tests when AgentCore is the answer, and this practice question on intelligent prompt routing covers sending each request to the cheapest model that can handle it.

AI safety, security and governance (20%)

Safety is a fifth of the exam, more than either of the last two domains. It covers input and output protection with Bedrock Guardrails, hallucination checks, prompt-injection detection and defence in depth; data privacy with VPC endpoints, Lake Formation, and PII detection with Comprehend and Macie; governance with model cards, data lineage and CloudTrail auditing; and responsible AI, including fairness evaluation and using a model as a judge.

Two details show up often. Guardrails can protect a model that doesn't run on Bedrock, as this practice question on applying a guardrail outside Bedrock shows. And an agent's limits belong in policy outside the prompt, which this practice question on authorizing agent tool calls is about.

Operational efficiency and optimization (12%)

This domain covers cost, speed and monitoring once the app is live. Cost means token efficiency, choosing a cheaper model tier where it is good enough, batching, and caching, including prompt caching. Performance means trading latency against cost, streaming, tuning retrieval, and inference settings such as temperature, top-k and top-p. Monitoring means CloudWatch metrics for tokens and quality, Bedrock invocation logs, tracking tool and agent calls, and keeping the vector store healthy.

This practice question on prompt caching covers the most common cost trick: a long document resent with every question.

Testing, validation and troubleshooting (11%)

The smallest domain is about proving an app works and fixing it when it doesn't. It covers Bedrock model evaluations, model-as-judge scoring, evaluating RAG pipelines and agents, and quality gates that stop a worse version from shipping. Troubleshooting covers context-window overflow, integration errors, prompts that stopped working, and retrieval that returns the wrong passages. Start with this practice question on evaluating a RAG application.

The services to know

AWS's in-scope list is long, and AWS calls it non-exhaustive. Group it like this:

Generative AI core: Amazon Bedrock, Bedrock AgentCore, Knowledge Bases, Prompt Management, Prompt Flows, Amazon Titan models, Amazon Quick, Amazon SageMaker AI, SageMaker JumpStart and the SageMaker Model Registry.
Data and search: Amazon OpenSearch Service, Aurora, DynamoDB, RDS, Neptune, DocumentDB, Kendra, AWS Glue and Athena.
Integration: Lambda, Step Functions, API Gateway, EventBridge, SQS, ECS, EKS, Fargate, Amplify and Cognito.
Also named: Kiro, Amazon Q Developer (whose IDE plugins AWS will stop supporting on April 30, 2027, pointing users to Kiro), Comprehend, Lex, Rekognition, Textract, Transcribe, Amazon Connect, Macie, KMS, WAF, X-Ray and CloudWatch.

Named in the guide, but closed to new customers

A few services in the in-scope list are closed to new customers as of October 2026: Amazon Q Business (folded into Amazon Quick), AWS App Runner, and the SageMaker AI features Augmented AI, Clarify, Ground Truth and Model Monitor. Two skills also mention Bedrock agent tracing and agent evaluations. The original Amazon Bedrock Agents is closed to new customers too, and AWS points new agent work to AgentCore. Learn what these did, because an answer option may name them, but study the current replacements (Amazon Quick, AgentCore) as the services AWS wants you to build on.

What's out of scope

The guide says the exam does not test model development and training, advanced machine learning techniques, or data and feature engineering. You are expected to use foundation models, not train them. Services listed out of scope include Amazon Redshift, AWS Batch, Elastic Beanstalk, Direct Connect, Transit Gateway, VPN and Amazon Forecast.

If you are unsure where fine-tuning ends and retrieval begins, the RAG vs fine-tuning guide covers that line, which the exam tests in the model selection tasks.

How AIP-C01 differs from AI Practitioner and ML Engineer

AWS has three exams in this area, and they are easy to confuse:

Level: AI Practitioner (AIF-C01) is foundational. Machine Learning Engineer – Associate (MLA-C02, in beta as of October 2026) is associate. Generative AI Developer (AIP-C01) is professional.
What you do: AI Practitioner asks you to understand AI and pick the right service, without building anything. ML Engineer asks you to prepare data and train, deploy and monitor models, mainly on SageMaker AI. Generative AI Developer asks you to build applications on foundation models that someone else trained.
Format and price: AI Practitioner is 65 questions in 90 minutes for 100 USD. The MLA-C02 beta is 85 questions in 170 minutes for 75 USD. AIP-C01 is 75 questions in 180 minutes for 300 USD.
Renewal: passing AIP-C01 also renews AI Practitioner, ML Engineer – Associate and Data Engineer – Associate, according to AWS's recertification page.

For the wider choice across vendors, see the best AI certifications guide. For the ML Engineer transition, see MLA-C01 vs MLA-C02.

Who should take it, and what to take first

AIP-C01 fits developers and engineers who already ship generative AI features on AWS: RAG chatbots, document processing, internal assistants, agents that call company APIs. If you've never built one, the scenarios will feel abstract, because each assumes you know what goes wrong in production.

AWS suggests, but does not require, one of these first: AI Practitioner, Solutions Architect – Associate, Machine Learning Engineer – Associate or Data Engineer – Associate. A sensible order:

AI Practitioner, if generative AI vocabulary is new to you. It is the cheapest way to learn what Bedrock, Knowledge Bases, Guardrails and agents are for, and AIP-C01 renews it later.
Solutions Architect – Associate, if you are new to AWS itself. AIP-C01 assumes you know VPCs, IAM, Lambda and API Gateway.
A project. Build one RAG application and one agent with a tool before you book. The exam's troubleshooting and optimization questions come from that experience.
AIP-C01.

The AWS certification roadmap guide shows where it fits among all of AWS's exams.

AIP-C01 cost and how to prepare

The exam costs 300 USD as of October 2026. AWS's exam-policy page lists local prices for each tier, and the 300 USD tier is INR 25,659, payable in rupees only through the Pearson voucher store. If you already hold an AWS certification, AWS gives you a 50% discount voucher that you can use on any other exam, which halves the price.

AWS's certification page links an exam prep plan on AWS Skill Builder: a practice question set and pretest, courses, Builder Labs, AWS Cloud Quest and AWS Jam, SimuLearn, and an official practice exam. The page doesn't say which of these are free, so check what your Skill Builder plan includes. AWS's separate certification prep page says its Official Practice Question Sets are free.

There is no AIP-C01 practice bank on this site yet. The AWS AI Practitioner practice tests cover the same services at a simpler level: Bedrock, Knowledge Bases, Guardrails, AgentCore, prompt caching, evaluation and routing. Use the full AI Practitioner question list to pick out the generative AI topics and check that the basics are solid before you move on to professional-level scenarios.

A practical plan:

Read the exam guide's task statements and mark every skill you've never done by hand.
Build in a sandbox account: a knowledge base over your own documents, a guardrail with denied topics and PII filters, an agent that calls one API, an evaluation job, and CloudWatch dashboards for tokens and latency.
Work through practice questions on each domain, starting with domain 1 and domain 3. Together they are just over half the exam.
Take the official practice exam under time: 75 questions in 180 minutes.

Frequently asked questions

Is the AWS Generative AI Developer Professional certification worth it?

It is worth it if you build generative AI applications on AWS, or want to. It is AWS's professional-level exam for generative AI, and it covers the work these jobs actually involve: retrieval, agents, safety, evaluation and cost. If you don't build software, AI Practitioner is the better fit.

How hard is AIP-C01?

It is a professional exam: 180 minutes, against 130 for most associate exams, with a higher passing score (750 versus 720) and long scenarios. AWS recommends two years of building production apps and one year of hands-on generative AI. Without that experience, the troubleshooting and optimization questions are the hardest part.

Do I need AI Practitioner before AIP-C01?

No. AWS has no prerequisites. It suggests AI Practitioner, Solutions Architect – Associate, ML Engineer – Associate or Data Engineer – Associate as useful first steps, and passing AIP-C01 later renews AI Practitioner.

How much does the AWS Generative AI Developer Professional exam cost?

300 USD as of October 2026, or INR 25,659 through the Pearson voucher store. With the 50% voucher from an earlier AWS certification, it costs half that.

Is the exam still in beta?

No. The beta ended on March 31, 2026, and the standard version has been open for registration since March 2026, refreshed to include Amazon Bedrock AgentCore.

Does AIP-C01 test training or fine-tuning models?

Only at the edges. The guide puts model development and training out of scope, but it does include deploying a LoRA fine-tuned model and choosing between retrieval and customization. You need to know when to fine-tune, not how to train.