MLA-C02 Study Guide 2026: AWS Machine Learning Engineer Associate (Updated Exam)

MLA-C02 is the updated AWS Certified Machine Learning Engineer – Associate exam. It replaced MLA-C01 for English-language candidates on September 29, 2026, and becomes generally available in all languages on January 14, 2027. It keeps MLA-C01's four-domain structure but adds generative AI to every domain: vector databases, RAG pipelines, Amazon Bedrock knowledge bases, agent deployment, LLM evaluation, and token cost monitoring.

Most MLA study guides you'll find were written for MLA-C01. They cover roughly 75% of the new exam and miss the part AWS changed most. This guide is built from the official MLA-C02 exam guide and AWS's own C01-to-C02 comparison.

MLA-C02 Key Dates: Beta, Retirement, and GA

MLA-C02 is in beta now, MLA-C01 is already gone in English, and GA lands on January 14, 2027. Here's the full timeline from AWS's September 2026 certification update:

If you were halfway through an MLA-C01 course in English, you can't take that exam anymore. Your options are the beta now or the GA exam in January. A C01 course still covers most of the material you need, so don't throw it out. Use it for Domain 1 and the classic SageMaker material, then add the generative AI content covered below.

MLA-C02 Exam Format, Cost, and Passing Score

The beta and GA versions differ in length and price. The scoring model is the same for both.

Question types are multiple choice (one correct answer out of four) and multiple response (two or more correct answers out of five or more, and you must select all of them). Scoring is compensatory, so a weak domain won't fail you if your overall score clears 720.

Beta results. AWS's certification page says beta results are "typically available within 5 business days of exam completion." That's much faster than older AWS betas, which made candidates wait up to 90 days. Check your score report timing when you book, since AWS can change it.

What Changed From MLA-C01 to MLA-C02

MLA-C02 keeps the four domains and their task statements but adds about 45 new skills, most of them about generative AI. Domain weights moved slightly:

The weights barely changed, but the content inside each domain changed a lot. From the official comparison page:

Added to MLA-C02

• Vector stores and embeddings: configure OpenSearch Service, RDS with pgvector, and S3 as vector databases; use embedding models on text and image data.
• RAG end to end: chunking and metadata extraction, choosing a RAG architecture, retrieval strategies and reranking, Bedrock knowledge bases, and RAG monitoring that includes retrieval accuracy.
• Foundation model selection and customization: choose FMs from Amazon Bedrock, fine-tune and distill them, prepare data for continued pre-training, and import externally built models with Bedrock Custom Model Import.
• Agents: deploy agents and configure their communication protocols, manage agent state, set up agentic workflow infrastructure, version agents in deployment pipelines, and detect coordination failures and tool failures.
• LLM evaluation: BLEU, ROUGE, BERTScore, semantic similarity, human-in-the-loop review, LLM-as-a-judge, and Amazon Bedrock evaluations.
• GenAI cost and security: token usage and embedding computation costs, vector storage optimization, choosing between Bedrock API keys and IAM credentials, and Amazon Bedrock Guardrails for sensitive data protection.
• Data hygiene for AI: masking, redaction, and anonymization; prompt-response pair validation; content safety screening.

Removed from MLA-C02

• SageMaker Neo and edge-device optimization
• Bring your own container (BYOC) with SageMaker
• Model size reduction (pruning, compression, changing data types)
• Loading training data from EFS/FSx
• Fine-tuning through SageMaker JumpStart custom datasets, which C02 rewrote as broader FM customization skills
• EventBridge infrastructure monitoring and provisioned-concurrency/quota troubleshooting

I'm glad Neo and BYOC are gone. Across the production AWS accounts I've worked in, edge compilation came up once and BYOC mostly showed up as a workaround someone wanted to delete. Most of the GenAI additions, on the other hand, are now everyday work on any team running Bedrock in production: chunking strategy, reranking, token spend, and guardrails. The exam is closer to the actual job than it was a month ago.

MLA-C02 Domains: What to Study

The official exam guide expects at least one year of experience with SageMaker AI and Amazon Bedrock, plus a year in a backend, DevOps, data engineering, or data science role. That "and" is new. Under C01, you could pass on SageMaker experience alone. Under C02 you can't.

Domain 1: Data Preparation for ML and AI (28%)

This is still the largest domain. The classic material carries over: ingesting and storing data, feature engineering, SageMaker Data Wrangler and Feature Store, Glue, bias metrics with SageMaker Clarify, and handling outliers and missing values. On top of that, you now need to know how to configure a vector store and prepare documents for RAG. Learn the tradeoffs between OpenSearch Service, RDS with pgvector, and S3 vector storage. Our vector database comparison explains why pgvector is usually the pragmatic default, and that reasoning carries directly into exam scenarios.

Domain 2: ML Model and FM Development (24%)

Expect "which approach fits" questions that weigh a custom model against a managed AI service, a pre-trained model, or a foundation model, balancing performance, latency, and cost. Know when prompt engineering is enough, when to fine-tune, and when RAG is the better fix. Know which evaluation metric fits which task: ROUGE for summarization, BLEU for translation, BERTScore or semantic similarity when wording varies, and LLM-as-a-judge when no reference answer exists. Hyperparameter tuning and the built-in SageMaker algorithms are still in scope.

Domain 3: Deployment and Orchestration (24%)

This domain gained the most weight and most of the agent content. SageMaker endpoint types (real-time, serverless, asynchronous, batch) and auto scaling still matter. New topics include FM hosting options, GPU scaling, Bedrock knowledge base setup, Bedrock Prompt Management, and CI/CD for agents and fine-tuned model versions. Amazon Bedrock AgentCore is explicitly listed as an in-scope service. If you haven't deployed an agent, work through our Amazon Bedrock AgentCore guide. It covers the runtime, memory, and gateway pieces that the "agent state management" and "agent communication protocols" skills point at.

You should be comfortable querying a knowledge base directly, since exam scenarios often ask why retrieval quality is poor:

aws bedrock-agent-runtime retrieve \
  --knowledge-base-id KB12345678 \
  --retrieval-query text="What is our refund window for annual plans?" \
  --retrieval-configuration '{"vectorSearchConfiguration":{"numberOfResults":5}}'

Look at the returned chunks and their scores. If the right passage isn't in the top 5, generation can't fix it. The likely causes are chunking, the embedding model, or a missing reranker, and C02 questions test exactly that line of diagnosis.

Domain 4: Operating, Monitoring, and Securing (24%)

Model Monitor, CloudWatch, IAM, KMS, and VPC endpoints carry over from C01. New topics: token and embedding cost monitoring, monitoring agent resource use, Bedrock evaluations for production FMs, scanning CI/CD pipelines with Amazon Inspector and CodeGuru, Bedrock API keys vs IAM credentials, and Bedrock Guardrails. Expect at least one scenario asking which credential type a workload should use. IAM roles are almost always the right answer for anything running inside AWS. Bedrock API keys exist for quick exploration and external tools, not for production services.

Should You Take the Beta or Wait for GA?

Take the beta if you already run both SageMaker and Bedrock workloads. At $75, it costs half of the GA price. You also get the credential three months before everyone else, and results come back fast. The costs are 20 extra questions, 40 extra minutes, and almost no third-party practice material yet. Pace yourself to about 2 minutes per question and flag anything that takes longer.

Wait for GA (January 14, 2027) if your Bedrock experience is mostly tutorials. The new content is spread across all four domains, so a candidate who is strong on SageMaker but light on generative AI can lose points in every domain. By January, practice exams will have caught up to the C02 blueprint.

For a 6–8 week plan, spend weeks 1–2 on Domain 1, including building a small RAG pipeline. Spend weeks 3–4 on Domain 2, comparing evaluation metrics against your own pipeline's output. Use weeks 5–6 to deploy one SageMaker endpoint and one Bedrock agent with a CI/CD pipeline. Save weeks 7–8 for monitoring, guardrails, cost, and timed practice. Build things rather than only reading about them. The C02 additions are almost all "configure and operate" skills, and they stick much better once you've seen a retrieval return the wrong chunks.

Where MLA-C02 Fits in the AWS AI Track

MLA-C02 now overlaps more with the AWS Generative AI Developer – Professional (AIP-C01), but the two still test different levels. MLA-C02 is an Associate exam that checks whether you can operate ML and FM workloads: prepare data, deploy, monitor, and secure them. AIP-C01 is a Professional exam ($300, 750 to pass) that goes deeper on designing generative AI applications. The MLA-C02 exam guide explicitly lists "designing and architecting full end-to-end AI and ML solutions" as out of scope.

If you're coming from the retired Machine Learning Specialty, MLA-C02 is now the closest replacement. See our MLS-C01 retirement guide for the full map. A sensible order for most engineers is AIF-C01 (optional), then MLA-C02, then AIP-C01.

Practice MLA-C02 Skills Hands-On With CloudaQube

Most of the new MLA-C02 content is about configuring and operating real resources: vector stores, knowledge bases, agents, guardrails, and cost monitoring. CloudaQube's AI & ML labs give you a real AWS environment to build RAG pipelines and Bedrock workloads step by step, without risking your own account or bill. Use them for weeks 1–6 of the plan above and you'll go into the exam having already done the tasks it asks about.