Model Selection

Claira routes each scan through an AI model you select for the case. The model choice affects latency, cost (tokens), and how the model handles difficult documents. This page explains where to change the model and how to validate the choice before scaling up.

Where to choose the model

Model selection is configured in the Claira Settings area for the active case.

1. Open your case workspace and select the Claira pane.
2. Open the Modes menu and choose Settings.
3. Locate the Model control (or equivalent model picker) and select the model your team should use for scans.

Available models

Claira offers region-resident model options plus a way to connect your own.

• Gemini Flash 3.5 (default for Canada tenants). The newest Gemini Flash generation, with a much larger context window than Gemini Flash 2.5 — useful for long documents that previously had to be truncated. Hosted inside the Canadian data residency boundary. Does not appear in the picker for Australia tenants, where Gemini Flash 2.5 remains the default.
• Gemini Flash 2.5 (default for Australia tenants). Lower latency and lower token cost. Best for high-volume scans, first-pass coding, and most everyday review tasks. Runs in your tenant's home region (Canada for CA tenants, Australia for AU tenants).
• Gemini Pro 2.5. A stronger model for complex, ambiguous, or borderline documents (for example, nuanced privilege calls or dense regulatory text). Slower per document and more tokens per scan than the Flash models, but better at deeper reasoning. Runs in your tenant's home region under the same residency guarantees as the Flash models.
• Claude Sonnet 4.6 (Australia tenants only). Anthropic's Claude model, hosted on AWS in Sydney inside the Australian data residency boundary. A strong alternative when the Gemini models are busy, and a second opinion from a different model family on borderline documents. Text scans only — media scans (image, audio, video) require a Gemini model. Does not appear in the picker for Canada tenants.
• Amazon Nova Lite (Canada tenants only). A fast, lightweight model hosted on AWS inside the Canadian data residency boundary. Useful for quick first-pass triage or as a second opinion from a different model family. Text scans only. Does not appear in the picker for Australia tenants.
• GPT-OSS 120B (Canada tenants only). An open-source model hosted inside the Canadian data residency boundary. Useful when your review protocol calls for an open-weights model alongside the Gemini options, or when you want a second opinion from a different model family on a benchmark set. Does not appear in the picker for Australia tenants.
• Bring Your Own Model. Connect your organization's own model deployment. Selecting this opens the setup documentation in a new tab.

Switching between the models above (where available) is a model-name change only — your prompt, fields, and case settings carry over unchanged. Your plan is charged the same per document whichever model you choose.

Models and Background Processing

Background Processing always runs on Google Gemini, regardless of the model selected in Settings. If a non-Google model (such as Claude Sonnet 4.6) is selected when you start a background task, the confirmation dialog shows a note that the background run may use a different model, and the task's history records the model that actually ran. Foreground bulk scans are unaffected — they use exactly the model you selected.

Before you change models mid-review

• Confirm field mappings still match your prompt output. The model does not change your fields, but different models may phrase answers slightly differently. Re-run a small benchmark set after switching.
• Re-test 10–25 documents with Single Review before running a new Bulk Scan.
• Check token usage in the Usage popover if your plan treats models differently.

Model settings (temperature, Top P, reasoning)

After you pick a model, you can tune behavior with Temperature, Top P, and Reasoning level. See Model Settings for defaults and task-based recommendations.

If results change after switching models

1. Run the same prompt on the same benchmark documents you used before the switch.
2. If outputs drift, tighten the prompt (definitions, allowed values, output format) before blaming the model.
3. If only one model fails with Error Codes such as E-AI, try the other available model and contact support if both fail.

Related

• Understanding Results
• Troubleshooting

Need help? Contact us at support@claira.to.