Inference level
The inference level controls which AI model Opal uses and how deeply the model reasons before responding. Higher inference levels pair more powerful models with deeper thinking to produce higher-quality output, but take longer and use more credits. Lower levels respond faster and use fewer credits, making them ideal for straightforward tasks.
Inference levels
Note
Opal uses Balanced by default. Balanced fits most tasks, adding reasoning to Standard while keeping responses fast. Choose a higher level for complex work, or a lower level for faster responses.
Inference level options include the following:
Inference level | Best for | Example user prompt |
|---|---|---|
Quick | Simple questions and quick tasks. | Give me a short headline variation for this landing page. |
Standard | Everyday tasks that don't need deep reasoning. | Summarize last week's web analytics report for the team newsletter. |
Balanced (default) | Everyday tasks that benefit from reasoning. | Summarize last week's web analytics reports and highlight the three biggest changes in user behavior for the team newsletter. |
Complex | Multi-step planning and analysis. | Build a step-by-step plan for A/B testing our checkout flow, including metrics to track and rollout strategy. |
Pro | Nuanced evaluation and strategy. | Evaluate whether to run personalization or multivariate testing for our product recommendation carousel, taking into account traffic volume, statistical significance, and long-term revenue impact. |
Code | Code writing, explanation, and debugging. | Write a JavaScript snippet to trigger an Optimizely experiment on product detail pages when users scroll 50% down. |
Note
Code uses the most powerful model configuration. While optimized for code tasks, it can also handle non-code tasks that need maximum reasoning depth.
Model assignments by inference level
When you select an inference level, Opal runs it on a matching large language model (LLM). You do not choose the model yourself. The following table shows which Google model and Claude model Opal maps to each inference level.
Inference level | Google model | Claude model |
|---|---|---|
Quick | gemini-3.5-flash-lite (LOW thinking) | claude-haiku-4-5 (LOW thinking) |
Standard | gemini-3.8-flash (no thinking) | claude-sonnet-4-6 (no thinking) |
Balanced | gemini-3.8-flash (LOW thinking) | claude-sonnet-4-6 (MEDIUM thinking) |
Complex | gemini-3.8-flash (MEDIUM thinking) | claude-sonnet-4-6 (HIGH thinking) |
Pro | gemini-3.1-pro (MEDIUM thinking) | claude-opus-5 (HIGH thinking) |
Code | gemini-3.1-pro (HIGH thinking) | claude-opus-5 (XHIGH thinking) |
The thinking level in the previous table controls how much the model reasons. Thinking lets the model reason through problems step-by-step before responding.
Set the inference level
Set the inference level in one of two places, depending on the scope you want. Administrators set an organization-wide default for Opal Chat. Agent creators set a level for each specialized agent that overrides the default.
Set the organization-wide default
The organization default applies to Opal Chat for everyone in your instance. You need administrator access to change it. For step-by-step instructions on setting the default, see Get started with Optimizely Opal for administrators.
The selected level becomes the default for Opal Chat across your organization. Specialized agents override this default with their own inference level.
Set the inference level for a specialized agent
When you create a specialized agent, set its inference level to match the tasks it handles. This level applies to that agent only and overrides the organization default.
Set the inference level after you choose the agent's model provider and configure its prompt and tools. For the full workflow, see Create a specialized agent. To match a level to your task, see the table in the Inference levels section.
Note
If you use Opti ID, administrators can turn off generative AI in the Opti ID Admin Center. See Turn generative AI off across Optimizely applicationsTurn generative AI off across Optimizely applications.