Ship a model picker with a safe default and legible alternatives

Make the default the safe choice

When a user leaves the model picker closed, the product has already chosen a model for them. The public release of GPT-5.6 was postponed after US government cybersecurity and national security reviews, with officials reportedly asking OpenAI to start with a small group of government-approved users.

A picker earns trust when the default is safe, alternatives are task-scoped, labels explain capability, and release notes set expectations. The default choice is the first trust signal, because it says what the product believes is safe enough for ordinary work.

In Microsoft 365 Copilot, OpenAI's GPT-5.6 family became available in July 2026. OpenAI announced GPT-5.6 as the preferred model powering Microsoft 365 Copilot. The preferred-model label is a default promise.

Microsoft positioned GPT-5.6 in July 2026 as supporting Word drafting, Excel analysis, PowerPoint presentations, Chat responses, and Cowork deliverables. The July 2026 update made OpenAI-operated Copilot models tenant-controlled, enabling eligible commercial tenants for all users from 24 July 2026 unless administrators selected No users. If your product has a comparable spread of tasks, choose the model you can defend across that spread.

The finished picker leaves the user with an answer that matches the product's stated safety and quality bar. That choice is documented, tested, and reversible without making the user feel punished for using it.

Scope alternatives to the task, not the model menu

Offer alternatives only where the user has a task that changes the right answer. A model picker that dumps every available model into every surface teaches users that the product does not know what it is doing.

In the same July 2026 Copilot update, users could select GPT-5.6 directly where model selection was available. Anthropic's Claude Fable 5 was added to Copilot Cowork in July 2026 for longer, more complex work. Anthropic's Claude Sonnet 5 was offered in Microsoft 365 Copilot in July 2026 for PowerPoint, Word, and Cowork.

Claude Fable 5 was off by default in Copilot Cowork when it rolled out in July 2026. A stronger model can exist without becoming the ambient choice.

The picker shows a small set of options tied to the current task, such as a faster draft model, a deeper analysis model, or a longer-work model. Each option appears only where the task makes it meaningful.

Where teams go wrong, they treat every model as a user-facing option. The picker becomes a catalog, and the user inherits the product's uncertainty. The fix is to ask what the user is trying to do, then expose only the models that change the answer for that task.

Label capability in the user's language

Write labels that say what the model does better, slower, cheaper, or more carefully. Do not ask users to infer capability from model names.

OpenAI's GPT-5.6 family offers Sol, Terra, and Luna as performance, speed, and cost options, plus a reasoning-effort control with an Ultra mode. A label that says faster for drafts is more useful than a version number. A label that says more effort for complex work tells the user what they are trading.

If a model is scoped to longer work, say that plainly. If a model has a data-handling difference, make that difference visible before the user relies on it.

Anthropic's Fable 5.1 normally uses a 30-day safety retention window, while qualifying enterprise customers can choose zero retention until Enterprise Frontier Safeguards arrives.

The user can choose between options without reading a model card, because the interface already states the tradeoff in one line. The label names the task, the speed, the cost, the effort, or the data behavior that changes.

Set expectations before the model changes

Write release notes that explain what changed, who is affected, and what the user should expect next. A model swap is a behavior change, so the note should name the default, the new option, and the reason for the change.

Microsoft's Sept. 1 Copilot update made Claude Fable 5.1 available to eligible Cowork and Copilot Studio users, and OpenAI's GPT-6 Astra was added on Sept. 4. A delay belongs in the release note. Users need to know when a model is being tested, restricted, or expanded.

In 2021, medical trust researchers characterized trust as confidence in a person's task competence, and distrust as suspicion that motives or information quality may be compromised. Put the change in the release note before the user discovers it in the output.

The release note says which surface changed, which model is now the default, which models are optional, and what users should expect. The note also states how to switch back if the new default does not fit the task.