Microsoft's chat model can now answer with an interactive chart. Your app still decides what it may render
GPT-chat-latest in Microsoft Foundry adds GenUI answers and interleaved-thinking output on the existing endpoint. The interface is now part of the response, and needs a rendering contract.
Listen to this article · 5 min
AI-generated narration of the full article.
A sales manager asks the internal assistant how the team’s pipeline looks this quarter. Last week the answer was three paragraphs. This week it is a chart with hover details, followed a few seconds later by a fuller analysis. Nobody on the application team shipped a release. The model changed.
That is the scenario Microsoft’s October 14 update to GPT-chat-latest in Foundry makes possible. The model brings GPT-6 family intelligence to conversational and agentic applications, with stronger reasoning, coding, search, multilingual behavior, personalization and task creation, and adds two capabilities that change the shape of an answer. Developers get the improvements “through the existing GPT-chat-latest in Microsoft Foundry, no need to redeploy a new model.” Microsoft lists pricing at $5 per million input tokens, $0.50 per million cached input tokens and $30 per million output tokens.
Two changes to what a response is
GenUI lets the model return visual or interactive answers. Microsoft’s examples are a structured weather card “that your application renders as an interactive widget” and an embedded sprint-progress chart the user can hover over. The phrasing matters: the model returns a structure, and the application renders it.
Interleaved-thinking output, which Microsoft says is sometimes called “double texting,” surfaces useful content early while work on the full answer continues. Asked to recommend a CRM, the model might first send a quick summary of three options, then a detailed comparison moments later.
The UI now needs a contract
When a model’s output was text, the application’s job was to display it safely. When the output is an interface description, the application decides what the model is allowed to build. That is a contract, and it should be written down:
- An allowlist of components. The application should render only components it knows, with schemas it validates: a chart, a table, a card. Anything outside the list falls back to text. Microsoft’s announcement does not describe arbitrary code running in the user’s interface, and enterprise applications should not allow it.
- Data provenance in the widget. A chart looks more authoritative than a sentence. Every rendered number should trace to a source the user can open, or the interface is a more persuasive way to be wrong.
- Accessibility and localization. Interactive answers must work with screen readers and in every language the application serves. A chart with no text alternative is a regression for some users.
- Actions behind confirmation. If a widget can trigger a task, such as creating a ticket, it is a transaction, and the user should confirm it as one.
Early answers need a status
Progressive answers are good for users who are waiting. They create a problem for systems that act on answers: which message is the answer? If an agent downstream, a log, or a person copying into an email takes the first quick summary, and the full analysis later contradicts it, the record holds the wrong version.
Applications should mark early content as provisional, keep the final response as the one of record, and log both. Microsoft found with its own support agent that honesty about what the agent knows drives adoption; showing which part of an answer is still in progress is part of that honesty.
“No need to redeploy” cuts both ways
Getting new capabilities without redeploying is convenient. It also means behavior under the same model name can change without a change on your side: answer format, length, cost per answer and now the presence of interactive output. For applications with regression tests, compliance reviews or users trained on a stable experience, that is a release you did not schedule.
Check in Microsoft’s documentation which version controls exist for GPT-chat-latest in your deployment, and treat a behavior change on a “latest” endpoint as a change event: rerun your evaluation set, check rendering paths and confirm cost per answer. A model that changes is a release, whether or not anyone pressed deploy.
The enterprise chatbot is becoming an interface generator. That is good news for users, as long as the application, not the model, stays the author of what the interface may do.