Each Atlas AI agent uses one language model for all its tasks. The model you choose affects how well the agent reasons through multi-step questions, how quickly it responds, and what it costs to run. Selecting the right model at the start makes it easier to meet your quality and performance targets without having to rebuild the agent.
Match model to task complexity
The main factor in choosing a model is how much reasoning the agent needs to do. An agent that queries the CDF knowledge graph, retrieves time series data, and analyzes the results in a single response needs a model that can reason well across multiple steps. Agents with simpler workflows, like document Q&A or single-tool lookups, can often perform well with a smaller, faster model.
| Task type | Model recommendation |
|---|
| Multi-step queries, multiple tools in sequence | Higher-capability model (for example, GPT-5 series, Claude 4.5 Sonnet, Gemini 2.5 Pro) |
| Single-tool lookups, document Q&A, or summarization | Mid-tier or smaller model (for example, GPT-5 mini, Claude 4.5 Haiku, Gemini 2.5 Flash) |
| High-volume agents where response speed matters | Smaller, faster model; verify quality meets your requirements with evaluation test cases |
Model availability by cloud
The models available to your agent depend on the cloud provider and region of your CDF project. Atlas AI supports models from Azure OpenAI, Google Cloud, and AWS.
See the Language model library for the full list of available models, their release dates, and their retirement dates.
Test models with evaluations
Start development with a higher-capability model to establish a quality baseline. Once the agent is working well, test whether a smaller model maintains acceptable quality for your use cases. Use your evaluation test cases to compare responses from different models before making a change.
Run your evaluation test cases after every model change, including automatic upgrades at retirement. A model change can affect responses in unexpected ways.
Model lifecycle
Cognite continuously updates the available models. Each stable model has a published retirement date. If you do not migrate your agent to a newer model before the retirement date, Atlas AI automatically upgrades the agent to the latest stable model.
To avoid unexpected behavior after an automatic upgrade, monitor the retirement dates in the Language model library and plan migrations before the deadline. After migrating, run your evaluation test cases to confirm the agent’s responses meet your expectations.
