AI-300 vs DP-100: what changed?
AI-300 replaced DP-100 when the Azure Data Scientist Associate certification retired on June 1, 2026. Much of the Azure Machine Learning content carries over — workspaces, compute, MLflow, pipelines, endpoints — but the exam now tests operating models rather than building them, and nearly half of it is about generative AI in Microsoft Foundry. If you studied for DP-100, you have a solid base for two domains and little for the other three.
The two exams side by side
| DP-100 (retired) | AI-300 | |
|---|---|---|
| Certification | Azure Data Scientist Associate | Machine Learning Operations Engineer Associate |
| Status | Retired June 1, 2026 | Current |
| Passing score | 700 | 700 |
| Focus | Design, train and deploy models | Automate, deploy, monitor and optimise ML and generative AI |
| Named DevOps tooling | Little | GitHub Actions, Bicep, Azure CLI |
| Languages offered | 10 | English only |
The shift in one sentence
DP-100 asked whether you could produce a good model. AI-300 asks whether you can put a model into production repeatably, roll it back safely, notice when it degrades, and do the same for a generative AI application whose “model” is a foundation model plus prompts plus retrieval.
Microsoft’s own comparison puts it the same way: training and evaluation stay, but automation, infrastructure as code, CI/CD, lifecycle governance, observability, drift detection and cost control move to the centre.
Domains compared
| DP-100 area | Where it went in AI-300 |
|---|---|
| Design and prepare a machine learning solution | Became Design and implement an MLOps infrastructure, now with Bicep, GitHub Actions and network isolation |
| Explore data and run experiments | Folded into Implement machine learning model lifecycle and operations; notebooks and AutoML remain |
| Train and deploy models | The core of the lifecycle domain, extended with rollout, rollback and drift monitoring |
| Optimize language models for AI applications | Expanded into three domains: GenAIOps infrastructure, quality and observability, and optimisation |
What carries over
If you prepared for DP-100, these parts transfer almost directly:
- Workspaces, datastores, compute targets and data assets.
- Environments and components, and building pipelines from components.
- MLflow experiment tracking and model registration.
- AutoML and hyperparameter sweeps.
- Managed online and batch endpoints.
- Responsible AI evaluation of a trained model.
Refresh them against the new outline rather than relearning them.
What is new or much deeper
Infrastructure as code. Deploying workspaces and their resources with Bicep and the Azure CLI, and running that deployment from a GitHub Actions workflow, are explicit objectives. So is configuring secure GitHub integration with Azure Machine Learning.
Network isolation and identity. Restricting network access to workspaces, and managed identities with role-based access control for both Azure Machine Learning and Foundry.
Registries. Sharing environments, components and models across workspaces, which is how a model moves from dev to prod without being retrained.
Safe deployment. Progressive rollout and rollback on endpoints, plus testing and troubleshooting a failing deployment.
Production monitoring. Data drift, production performance metrics, and triggering alerts or retraining when a threshold is crossed.
GenAIOps. Foundry resources and projects, foundation model deployment options including provisioned throughput units, and prompts versioned in Git.
Generative AI evaluation and observability. Quality metrics such as groundedness and relevance, risk and safety evaluations, automated evaluation workflows, tracing, and token cost tracking.
RAG and fine-tuning optimisation. Chunk size, similarity thresholds, hybrid search, embedding model choice, synthetic data and advanced fine-tuning methods.
What faded
DP-100 spent more time on data exploration and on the mechanics of choosing an algorithm. The AI-300 outline still names notebooks and AutoML, but only as tools for finding a model to operationalise. Detailed feature engineering and training-script authoring are no longer where the marks are. Microsoft notes that related topics may still appear, so treat this as a shift in weight, not a hard boundary.
Is it a harder exam?
For most DP-100 holders, yes. The machine learning half is familiar, but the DevOps tooling and the generative AI half are new, and together they are the larger part of the exam. See is AI-300 hard? for the detail.
If you already hold DP-100
The credential stays on your Microsoft Learn profile, but it cannot be renewed: the renewal assessment was retired with the certification. If it matters for your job or your employer’s partner status, plan for AI-300.
The shortest route is the four-week study plan, moving quickly through week two (model lifecycle) and spending the saved time on weeks one and three (infrastructure as code and Foundry).