ChatGPT can now return actual interactive UI elements, not just text. Rolled out in early October 2026, Intelligent UI lets model responses include interactive visuals and task-specific mini‑apps; OpenAI also published a beta Decisions API and a lower-latency “ultrafast” response mode exposed as a separate response tier in the Responses API.
This is the single most consequential change here: responses are now a runtime, not only data. That sounds like a UX win — and it is — but it also moves a new execution surface into the middle of conversational flows. Platform teams need to treat those UI payloads like code: signed, sandboxed, versioned, and auditable.
What actually shipped
- Rollout: these capabilities began appearing in ChatGPT in early October 2026. OpenAI's documentation lists Intelligent UI and a Decisions API beta; availability varies by account and product tier.
- Intelligent UI: model responses can include a machine-readable UI manifest or component specification that instructs clients how to render interactive elements. Examples include forms, buttons that trigger follow-up actions, small data tables, and embedded controls generated by the model rather than hand-crafted HTML.
- Decisions API (beta): returns structured, typed answers (from text and images) with lower end-to-end latency than token-streaming responses, targeting programmatic decisioning and deterministic workflows.
- Ultrafast mode: OpenAI exposed a lower-latency response tier to reduce inter-token delays and speed up streaming output for latency-sensitive flows.
A few technical implications
First, Intelligent UI flips the trust boundary. Previously, you controlled the UI layer and called the model for content. Now the model can propose UI elements; unless the platform intercepts and validates them, you are effectively rendering model-generated client code. That creates obvious attack vectors: prompt-poisoned UI that exfiltrates tokens or persuades users to perform privileged actions, data leakage through unconstrained form submissions, and subtle UX manipulations.
Second, the Decisions API is an API-design shift. It trades token streaming for typed, low-latency answers — good for programmatic decisioning and deterministic workflows. Teams building orchestrators or agents will value the latency and structured output, but the risk is architectural: you'll be tempted to push business logic into LLM decisions that are harder to version, test, and observe than service code.
Third, availability and tiering matter operationally. Paid and free accounts may be routed to different model variants or configurations, creating divergent performance, cost, and availability profiles. Choosing which variant to base automation or critical decision paths on is now an explicit platform decision.
What to do now (hard take)
Treat Intelligent UI outputs like executable artifacts. Force model-generated interfaces through a rendering pipeline that:
- verifies a signed manifest, or at least enforces an allowlist of UI primitives and disallows arbitrary script execution;
- runs in a sandboxed iframe with strict CSP and no access to credentials or host cookies;
- records an immutable audit trail for who saw what UI and what actions were taken.
Also: use the Decisions API only where you can accept eventual model drift. If a call affects billing, payroll, or security posture, implement deterministic server-side guards and explicit human-in-the-loop confirmations.
This is the right product move from OpenAI — lower-latency, structured decision outputs and richer UX accelerate real workflows — but it's overdue for better enterprise controls. Platform teams that treat these as drop-in text APIs will get burned.
Want a concrete next step? Add "model-generated UI" to your threat model and CI gating today. Expect vendors and reverse-proxy providers to ship signed-tool registries, CSP tooling, and UI-manifest validators within months; until then, assume any model-provided interface is untrusted.
For a comparison on agent governance and model-integration patterns, look at vendor work such as Amazon's Bedrock discussions around agent governance and EKS integration.
Prediction: within a year enterprise controls will require signed UI manifests and session-bound tokens for any model-generated UI that can trigger actions. If your platform doesn't plan for that now, you'll be retrofitting brittle bolt-ons while the rest of the stack moves on.