Models, modes, and privacy
Bergur DavidsenUpdated 2026-07-14
The model and mode shape how Usable Chat responds, how much reasoning effort it requests, which tools are compatible, and how a request may be routed. The available choices depend on deployment, account access, subscription, provider configuration, and embed restrictions.
Choose based on the work and its sensitivity, not only on the largest model name.
Curated model selection
The Curated picker presents reviewed choices intended for common tasks. Use it when you want a sensible speed, cost, and capability balance without configuring routing details.
A fast general model is usually sufficient for summarization, extraction, drafting, and straightforward retrieval. Choose a heavier reasoning option for difficult comparisons, multi-step analysis, or tasks where mistakes are expensive enough to justify slower work.
Model names and tiers change over time. Read the current picker rather than treating an old documentation example as a permanent inventory.
Advanced selection
Where enabled, the Advanced picker lets users choose a registry model and explicit modifiers. Released controls can include:
- no, low, medium, or high requested reasoning effort;
- a Zero Data Retention requirement where supported;
- a non-ZDR route with a visible warning;
- an OpenRouter-only route preference.
The model tooltip or accessible label can expose the resolved base model, reasoning setting, privacy requirement, and route. Review those values before sending sensitive work.
A route preference is not a permission bypass. Account access, model availability, workspace permissions, tool compatibility, and configuration still apply.
Reasoning and thinking displays
Reasoning modes request different effort from compatible models. More effort can improve difficult work, but it also increases latency and may increase usage.
A visible thinking or status panel is a progress surface, not a guarantee that private model reasoning will be exposed. Provider reasoning-only content is not treated as the final answer. Judge the result by its visible answer and sources.
Use deeper reasoning when the task requires it. Do not enable it automatically for routine lookups that should be answered directly from a source.
Direct routes and fallbacks
Some models can use a direct provider route, while others use OpenRouter. Released Chat behavior can retry certain direct Claude or Gemini failures through OpenRouter and show a visible retry status.
Fallback improves reliability, but it may change the route used for the request. If routing policy matters, inspect the selected route and privacy indicators before sending the prompt.
Zero Data Retention
A ZDR selection asks routing to use providers that support the required retention policy. Availability depends on the selected model and provider route.
If a non-ZDR warning appears:
- switch to a curated/private option;
- select a model and route that support ZDR;
- remove sensitive information;
- stop and confirm your organization's policy before continuing.
Do not interpret a model's general reputation as a retention guarantee. Use the current route indicator.
Locked and hidden choices
An Expert or embed configuration can:
- set a default model;
- restrict the allowed models;
- lock model choice;
- hide the picker;
- lock a mode or feature.
When the picker is missing, the assistant may be intentionally configured for one approved model. A hidden control is not evidence that the account can use every model.
Models and tools
Not every model supports every tool flow equally. Vision requires image-capable handling, and tool schemas must satisfy provider requirements. A tool-enabled answer can also fail even when ordinary text chat works.
When troubleshooting, record:
- the displayed model and route;
- reasoning and privacy modifiers;
- whether tools were enabled;
- whether the failure happened before, during, or after a tool call.
Do not post prompts, tokens, workspace IDs, or private source content in public support threads.
Choose safely
- Use the simplest capable model.
- Review route and privacy indicators for sensitive work.
- Keep confidential data inside approved workspace and provider boundaries.
- Prefer grounded source verification over model confidence.
- Treat generated legal, financial, medical, security, and operational guidance as material requiring human review.
- Re-check the picker after a release instead of hard-coding a model name into user guidance.
Troubleshooting
A model is unavailable
The account, subscription, provider, region, or embed allowlist may not expose it. Choose an available model or ask the configuration owner.
Reasoning behaves differently than expected
Check whether the selected model string includes an explicit reasoning choice. A bare model selection can use its configured default.
A provider request fails
Retry once, choose another route or model, and simplify the tool set. Do not repeatedly submit sensitive work while routing is unclear.
The model picker is hidden
The current Expert or embed may lock the model. Review the configuration rather than trying to override it from the browser.