ChatGPT's Intelligent UI, answers you can use

A new way to answer

OpenAI released GPT-6 in ChatGPT on 7 October 2026, together with a feature it calls Intelligent UI.¹¹1OpenAI, "GPT-6 and Intelligent UI for everyone", 7 October 2026. No date appears in the page text as read; the date comes from the release-notes entry of 7 October that links to it, and from the system card published the same day. Supports: the definition of Intelligent UI and the training to compose with "text, visuals and interactive elements"; more than 1.2 billion weekly users; the list of elements; the road trip, Sunday roast, savings calculator, bill splitter and game examples; the lamb roast demo with the GPT-5.6 Instant table (4 guests 3 to 4 lb of lamb and 2 lb of potatoes, up to 10 guests 6 to 7½ lb and 5 lb), the meal at 6:00 pm and the follow-up for 5 guests; the library and the compiler; training and evaluation of interfaces; "work ahead"; interleaved thinking and answering; 44% sooner for questions that need web search; the Extra High and Medium comparison; web search decisions; safety claims; Sol for paid tiers and Luna for Free and Go; rollout from 7 October and the next day; Enterprise admin settings; Work and Codex unchanged; the closing claim. https://openai.com/index/gpt-6-for-everyone/ ²²2OpenAI Help Center, "ChatGPT release notes", entry of 7 October 2026, read on 17 October 2026: the range "from side-by-side comparisons and diagrams to tools"; calculators, bill splitters and games; the automatic choice of format; answering while thinking or using tools; availability from Instant through Extra High; the Pro reasoning option on GPT-6 Astra; "workspace admin settings"; Work and Codex unchanged. https://help.openai.com/en/articles/6825453-chatgpt-release-notes The launch post describes it as "a new capability that allows ChatGPT to answer quickly with fully interactive user interfaces". According to the same post, those interfaces make "everyday answers more visual" and make learning complex topics easier. They can also "quickly create a tool to solve a task right in the moment".¹¹

OpenAI says it trained GPT-6 to compose responses "using text, visuals and interactive elements". The model chooses "how they fit together based on your question".¹¹ The help center gives a shorter definition. In its words, Intelligent UI lets ChatGPT "shape each answer" around what the user wants to understand or do, "directly in the conversation".³³3OpenAI Help Center, "Intelligent UI in ChatGPT", undated, read on 17 October 2026: the one-line definition; the list of elements; the choice of format; checklists and their state within and across threads; not in the Work tab; not in Voice; the Pro reasoning option; Sol and Luna both supporting Intelligent UI; reasoning levels; the Simple setting and its web path; custom instructions; per-conversation requests; connected plugins. https://help.openai.com/en/articles/20001598-intelligent-ui-in-chatgpt

The release notes of the same day describe the range of formats. It runs "from side-by-side comparisons and diagrams to tools you can use directly in the conversation".²² Calculators and bill splitters are among the tools the notes name.²²

GPT-6 is "built for more than 1.2 billion people who use ChatGPT each week", according to the launch post.¹¹ In its earlier post introducing apps in ChatGPT, OpenAI gave a reach of "over 800 million ChatGPT users".⁴⁴4OpenAI, "Introducing apps in ChatGPT and the new Apps SDK", date not visible in the text read on 17 October 2026; the page carries "Update on November 13, 2025": a reach of "over 800 million ChatGPT users", without the word weekly in the line read; an SDK that "extends MCP so developers can design both the logic and interface of their apps". https://openai.com/index/introducing-apps-in-chatgpt/ If both figures count weekly users, the difference is 400 million, a rise of 50%.⁵⁵5Arithmetic on two OpenAI statements: 1.2 billion minus 800 million is 400 million, and 1.2 divided by 0.8 is 1.5, a rise of 50%. The comparison holds only if both figures count weekly users. https://openai.com/index/gpt-6-for-everyone/ and https://openai.com/index/introducing-apps-in-chatgpt/ The line read in the apps post does not use the word weekly.⁴⁴

The launch post sets the release against "fitting every question into the same conversational format".¹¹ Its closing line reads "Instead of people adapting to software, software will adapt to people".¹¹ Elsewhere in the post, OpenAI calls Intelligent UI "a step toward a ChatGPT that can shape an interface around what you're trying to do".¹¹

Text, visuals and controls

A single list in OpenAI's pages sets out what a response can now include. It opens with "graphics" and "tappable buttons". Forms and charts come next, followed by "interactive experiences you can use directly in your conversation".¹¹³³ OpenAI does not present the list as complete. No full catalogue of the components appears on its pages.⁶⁶6Not stated on any OpenAI page read on 17 October 2026: a full catalogue of the components; whether interactive components run in a sandbox, open links or reach the network; what happens to data entered into forms; third-party code in Intelligent UI; latency in seconds or milliseconds; Intelligent UI in the API or the Apps SDK; any extra price or separate usage limit. Pages read: the launch post, the help center article on Intelligent UI, the release notes, the system card, the API model pages for GPT-6 Sol and GPT-6 Luna, and the apps post. https://openai.com/index/gpt-6-for-everyone/ and https://help.openai.com/en/articles/20001598-intelligent-ui-in-chatgpt

The launch post and the help center use the same lines to explain the choice of format. "The format depends on what you're asking. A comparison might work best side by side."¹¹³³ An explanation "might be easier to understand through an interactive diagram". When a plain answer is the most useful, OpenAI says ChatGPT can give that instead.¹¹³³ Marcus Mendes of 9to5Mac quoted the first of those lines in his report on the launch.⁷⁷7Marcus Mendes, 9to5Mac, "OpenAI brings GPT-6 to ChatGPT and debuts Intelligent UI", 7 October 2026: quotes "The format depends on what you're asking"; writes that interactive widgets let users explore concepts and change inputs to see how results shift. https://9to5mac.com/?p=1073252

Other examples in the launch post place the interface next to ordinary content. The stops on a road trip "can appear on a map, with notes on what's worth a detour". The timing for a Sunday roast "can sit alongside the recipe".¹¹

A user can also ask for a tool. OpenAI's examples include "a calculator to explore how savings could grow" and "a bill splitter for dinner with friends". A game played inside the conversation is another.¹¹ For learning, visual explanations let a user "change an input to see what happens". They also let a user "work through a problem step by step".¹¹ In his report for 9to5Mac, Mendes wrote that interactive widgets let users explore concepts in complex prompts, changing inputs to see how the results shift.⁷⁷

The help center names checklists as one kind of component.³³

The launch post shows the change with a demo in two columns. The prompt reads "Give me a plan for a sunday lamb roast, I'm having friends over still figuring out numbers tbh".¹¹ One column holds the GPT-5.6 Instant answer, set in markdown. It includes a menu.¹¹

It also has a table of lamb and potatoes for 4 to 10 guests. For 4 guests it gives 3 to 4 lb of lamb and 2 lb of potatoes. For 10 guests it gives 6 to 7½ lb of lamb and 5 lb of potatoes.¹¹ A schedule is timed for eating at 6:00 pm. One of the suggested follow-ups reads "Help me adjust the lamb roast plan for 5 guests".¹¹

The other column is labelled "GPT-6 Instant". Its version is visual, and its contents did not come through when the page was read as text.¹¹ Instant in that label is a reasoning level. The help center says Intelligent UI is available "from Instant through Extra High reasoning", and the models behind GPT-6 in ChatGPT are named Sol and Luna.³³¹¹

A library and a compiler

Behind Intelligent UI sits "a library of native, streamable components". OpenAI paired it with "a compiler that processes the interface as the model generates it".¹¹

The library gives each response "a familiar design foundation", according to the launch post. It leaves the model "room to decide how the pieces fit together".¹¹

The compiler lets the interface "appear progressively as the model generates it". It does so "without waiting for the entire response to be complete".¹¹

OpenAI says it gave GPT-6 "the freedom to create the most useful interface for each question". The same sentence sets a limit on that freedom. The experience was to stay "fast and familiar across web and mobile".¹¹

Earlier interfaces from OpenAI and Anthropic worked in a separate space. Anthropic's Artifacts "appear in a dedicated window alongside their conversation".⁸⁸8Anthropic, "Claude 3.5 Sonnet", page metadata 21 June 2024, 03:28 UTC, which is the evening of 20 June 2024 in US time: Artifacts "appear in a dedicated window alongside their conversation". https://www.anthropic.com/news/claude-3-5-sonnet OpenAI's canvas was "a new interface for working with ChatGPT on writing and coding projects". It "opens in a separate window" and was "built with GPT-4o".⁹⁹9OpenAI, "Introducing canvas", date not visible in the text read on 17 October 2026: "a new interface for working with ChatGPT on writing and coding projects"; "opens in a separate window"; "built with GPT-4o"; opens automatically "when ChatGPT detects a scenario in which it could be helpful". https://openai.com/index/introducing-canvas/ Canvas opened by itself "when ChatGPT detects a scenario in which it could be helpful".⁹⁹ Intelligent UI places its elements inside the answer, in the conversation itself.³³

Canvas has since been withdrawn from current models. From 28 May 2026 it was no longer available in GPT-5.5 Instant or GPT-5.5 Thinking.¹⁰¹⁰10OpenAI Help Center, "ChatGPT release notes", entries of 30 January 2026 ("More visual responses"), 10 March 2026 ("Interactive learning for math and science", "70+ math and science topics"), 28 May 2026 (canvas no longer available in GPT-5.5 Instant or GPT-5.5 Thinking, replaced by writing blocks and code blocks) and 22 September 2026 (interactive flashcards for "all ChatGPT plans"), read on 17 October 2026. https://help.openai.com/en/articles/6825453-chatgpt-release-notes The release note of that day says writing and coding are "now supported directly in chat responses through writing blocks and code blocks".¹⁰¹⁰

Apps in ChatGPT took another route. Their SDK is built on MCP, the Model Context Protocol. It "extends MCP so developers can design both the logic and interface of their apps".⁴⁴ In that design a developer draws the interface. In Intelligent UI the model composes it from OpenAI's library.¹¹

Learning when text is enough

OpenAI says it "expanded" its training methods for GPT-6. The aim was to help the model make "thoughtful decisions" about content and layout. Visuals and interaction were part of the same training.¹¹ The interfaces the model created were evaluated for clarity and usefulness. They were also judged on completeness.¹¹

From this training, the launch post says, the model learned "how to organize information clearly" and "when to use interactivity". It also learned "when a simple text response is enough".¹¹

The release notes make the same point more briefly. ChatGPT "can choose the format automatically, and still use a simple text answer when that works best".²² A user does not have to ask for an interface to get one. The model decides, and the user can steer the decision through settings.³³

OpenAI does not describe its design judgment as finished. "There's still work ahead to improve the model's design judgment and expand what it can create", the launch post says.¹¹

Visual answers in ChatGPT did not begin with GPT-6. A release note of 30 January 2026, titled "More visual responses", said answers "may include at-a-glance visuals".¹⁰¹⁰ Checking a team's latest stats was one example. Converting units was another.¹⁰¹⁰

On 10 March 2026 ChatGPT added "Interactive learning for math and science". It came as an "interactive visual module" covering "70+ math and science topics".¹⁰¹⁰ The Pythagorean theorem was one of them. The ideal gas law and lens equations were others.¹⁰¹⁰

On 22 September 2026 OpenAI released interactive flashcards for "all ChatGPT plans".¹⁰¹⁰ Intelligent UI followed 15 days later, on 7 October.²²¹¹¹¹11Arithmetic: from 22 September to 7 October 2026 is 15 days. https://help.openai.com/en/articles/6825453-chatgpt-release-notes

Answering while it thinks

GPT-6 also changes when an answer starts. OpenAI says it is making it possible for ChatGPT "to begin answering while it continues to think".¹¹ The launch post recalls the cost of reasoning models. With them, users "had to wait for the model to finish thinking before getting an answer". Now, it says, "ChatGPT can interleave thinking with answering."¹¹

The model was trained "to take into account the user's waiting time". It composes an answer "using its knowledge and findings up to that point".¹¹ OpenAI says it learned to build answers "across multiple partial responses". Each part is meant to add "useful information without unnecessary filler". The whole is meant to stay "as cohesive and factual as one written all at once".¹¹

The release notes add tools to the description. GPT-6 can begin answering "while it continues to think or use tools". It can "then add more findings without another prompt".²² In the notes' own words, "You can start reading before the full answer is ready."²²

OpenAI gives one figure for the gain. "For questions that need web search, GPT-6 Instant starts answering 44% sooner, on average, than GPT-5.6 Instant."¹¹ A wait 44% shorter is 56% of the old wait.¹²¹²12Arithmetic: 1 minus 0.44 is 0.56, so a wait 44% shorter is 56% of the old one. 10 seconds times 0.56 is 5.6 seconds. The 10 seconds are illustrative; OpenAI gives no absolute times. https://openai.com/index/gpt-6-for-everyone/ On that ratio, a wait of 10 seconds would fall to about 5.6 seconds. The 10 seconds are illustrative, since OpenAI publishes no absolute times.¹²¹²⁶⁶

Another comparison comes from an internal evaluation of "high-value, everyday agentic tasks". There, GPT-6 Extra High began answering in the same time as GPT-5.6 Medium.¹¹ Its overall score was better than that of GPT-5.6 Extra High. The post gives neither the times nor the scores.¹¹

OpenAI also says GPT-6 "makes better decisions about when to look something up". In an internal evaluation of difficult problems, it says the model "correctly addressed the key aspect of the user's question more often than GPT-5.6".¹¹ No number comes with that claim.¹¹

The interface streams in the same way as the text. The compiler shows it as "the model generates it", so components can appear while the answer is still growing.¹¹

Michelle Hawley of VKTR wrote that OpenAI's performance figures, including the 44% figure, "have not been independently verified".¹³¹³13Michelle Hawley, VKTR, "OpenAI Brings GPT-6 to All ChatGPT Users With Interactive Intelligent UI", 7 October 2026: OpenAI's performance figures, including the 44% figure, "have not been independently verified"; calculators and dashboards generated on the fly "can look authoritative"; admin control gives IT teams control over timing. https://www.vktr.com/ai-platforms/openai-brings-gpt6-to-all-chatgpt-users-with-interactive-intelligent-ui/

Which model, which plan

GPT-6 Sol powers GPT-6 in ChatGPT on the paid tiers, from Plus up to Enterprise. GPT-6 Luna serves the Free and Go tiers.¹¹ OpenAI says both are "tuned for everyday conversation".¹¹ The help center says both "support Intelligent UI".³³

The paid tiers started to receive GPT-6 with Intelligent UI on 7 October, in the Chat tab.¹¹²² Free and Go followed on 8 October, 1 day later.¹¹¹⁴¹⁴14Arithmetic: from 7 October to 8 October 2026 is 1 day. https://help.openai.com/en/articles/6825453-chatgpt-release-notes OpenAI says the rollout is global, and it states no regional exclusions.¹¹²²

Intelligent UI works at every reasoning level "from Instant through Extra High".²²³³ The Pro reasoning option is the exception. It "uses GPT-6 Astra and does not support Intelligent UI".³³²² Astra was the first GPT-6 model. It arrived on 3 September 2026, when access was "rolling out to a limited set of organizations".¹⁵¹⁵15OpenAI Help Center, "ChatGPT release notes", entries of 3 September 2026 (GPT-6 Astra, access "rolling out to a limited set of organizations"), 22 September 2026 (GPT-6 Sol and GPT-6 Luna in ChatGPT Work and Codex, "separate from the models available in Chat") and 29 September 2026 (GPT-6.1 Sol rolling out in Work and Codex, Pro first, outside Chat), and the entry on GPT-Live-1 showing "visual results through supported widgets", read on 17 October 2026. https://help.openai.com/en/articles/6825453-chatgpt-release-notes

Intelligent UI is limited to Chat. To the question of the Work tab, the help center answers "No". It gives the same answer for Voice.³³ The release notes describe GPT-Live-1, which can "show visual results through supported widgets".¹⁵¹⁵ That is a different feature, and Intelligent UI remains unavailable in Voice.³³

For Enterprise, "availability depends on workplace admin settings", in the launch post's phrase. The release notes say "workspace admin settings".¹¹²² The help center page on Enterprise workspace settings describes model controls for Chat. Admins "can choose an admin default model and reasoning level". That page names no toggle for Intelligent UI.¹⁶¹⁶16OpenAI Help Center, "Managing workspace settings in ChatGPT Enterprise", undated, read on 17 October 2026: admins "can choose an admin default model and reasoning level or allow User's last choice"; no Intelligent UI toggle named. https://help.openai.com/en/articles/8411955 Hawley wrote for VKTR that admin control gives IT teams control over the timing.¹³¹³

Work and Codex keep their models. "The models powering Work and Codex are not changing as part of this release," OpenAI says.¹¹²² Models named GPT-6 Sol and GPT-6 Luna reached Work and Codex earlier, on 22 September 2026. The release note of that day says they "are separate from the models available in Chat".¹⁵¹⁵

The system card for the new models settles the naming by month. October marks "the versions released today". September marks "the versions that remain in use in Codex and Work".¹⁷¹⁷17OpenAI Deployment Safety Hub, "GPT-6 Sol and GPT-6 Luna: October 2026 update", published 7 October 2026: naming by month of release; replacement of GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT; contents with no Intelligent UI section; High capability in cybersecurity and in biological and chemical domains, below High in AI self-improvement; evaluation settings; instruction hierarchy 99.99% and 99.79%; indirect prompt injection with GPT-RED 97.13% and 95.80%; regressions and "borderline but still generally safe"; Table 1, self-harm safe-completion rates 0.934 and 0.896, 0.932 and 0.901; factuality. https://deploymentsafety.openai.com/gpt-6-october The October versions of Sol and Luna power Chat. They "will replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT".¹⁷¹⁷

A third model sits outside Chat as well. Since 29 September 2026, GPT-6.1 Sol has been rolling out in Work and Codex, starting with Pro.¹⁵¹⁵ It is not part of Chat.¹⁵¹⁵

What the user controls

The help center takes up the question of turning Intelligent UI off. Its answer is a setting called Simple.³³ On the web, the path is Personalization, then Layout and Visuals, then Simple. The help center gives that path for the web only.³³

Simple "reduces visual and interactive elements, though some may still appear".³³ There is no setting that removes them all.³³

Custom Instructions can steer the format. The help center's examples are "keep answers concise" and "use tables when comparing options".³³ Such preferences "guide the response; the exact layout can vary".³³ A user "can also ask for a different style in an individual conversation".³³

Some components remember what the user did. Checklists are the example the help center gives. They "can retain their state when you refresh the chat thread".³³ That memory has a limit. "State is not retained across chat threads."³³

Plugins the user has connected "will continue to work in ChatGPT", in the help center's words. Gmail is one of its examples. Health and Finance are two more.³³ The help center does not say whether plugin results are drawn with Intelligent UI.³³

Writing for VKTR, Hawley raised a concern about tools built on the spot. Calculators and dashboards generated on the fly "can look authoritative", she wrote, and that raises questions about how users check their results.¹³¹³

What the system card covers

OpenAI published a system card for the October versions of Sol and Luna on 7 October 2026.¹⁷¹⁷ Its contents run from safe completions and vision to preparedness and safeguards. It has no section on Intelligent UI, and no entry on the interface or its components.¹⁷¹⁷

According to the launch post, GPT-6 builds "on several of Astra's safety advances". Relative to GPT-5.6 Sol, it is meant to resist attempts to bypass its safety training better.¹¹ OpenAI reports that in adversarial testing GPT-6 resisted such attempts more strongly, "particularly attacks that adapt across multiple turns".¹¹ The post also calls GPT-6 "better at recognizing and communicating the limits of its own capabilities".¹¹

In its preparedness assessment, OpenAI treats the October Sol and Luna "as High capability" in cybersecurity. The same rating applies in the biological and chemical domains. Neither model reaches the High threshold in AI self-improvement.¹⁷¹⁷

The card's safety evaluations ran "at their lowest reasoning deployment settings". Its capability assessments ran "at their maximum reasoning effort".¹⁷¹⁷

On instruction hierarchy, the card's measure of direct prompt injection, Sol scored 99.99% and Luna 99.79%.¹⁷¹⁷ Indirect prompt injection was tested with GPT-RED and averaged over reasoning levels. On that test Sol scored 97.13% and Luna 95.80%.¹⁷¹⁷

The card admits regressions. GPT-6 Sol (October) shows "a statistically significant regression on standard self-harm".¹⁷¹⁷ GPT-6 Luna (October) regresses on standard self-harm and on gore. It regresses on sexual content as well.¹⁷¹⁷ OpenAI judged the violations "borderline but still generally safe".¹⁷¹⁷

In the card's production benchmarks, Sol's safe-completion rate on standard self-harm fell from 0.934 for GPT-5.6 Sol (August) to 0.896.¹⁷¹⁷ Luna's fell from 0.932 for GPT-5.6 Luna (August) to 0.901.¹⁷¹⁷ The drops are 0.038 for Sol and 0.031 for Luna.¹⁸¹⁸18Arithmetic: 0.934 minus 0.896 is 0.038, and 0.932 minus 0.901 is 0.031. https://deploymentsafety.openai.com/gpt-6-october

On hallucinations, both October models "deliver significant improvements in factuality" over the August versions of GPT-5.6.¹⁷¹⁷ The gains hold "across nearly all evaluation metrics". The exception is "claim-level factual errors on factual-heavy prompts".¹⁷¹⁷ The card shows those results in a chart.¹⁷¹⁷

Before Intelligent UI

Several companies had tried interfaces generated by a model before October 2026.⁸⁸¹⁹¹⁹19Google Research, "Generative UI: A rich, custom, visual interactive user experience for any prompt", 18 November 2025: dynamic view and visual layout as "two experiments" in the Gemini app, and AI Mode in Search; "Gemini designs and codes a fully customized interactive response for each prompt"; "a minute or more"; "occasional inaccuracies in the outputs"; the preference evaluation, which "did not take into account generation speed". https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/

Anthropic's "Imagine with Claude" left the whole interface to the model. "No functionality is predetermined; no code is prewritten," Anthropic wrote.²⁰ It was a temporary research preview.²⁰

Google's generative UI took a similar line. In dynamic view, "Gemini designs and codes a fully customized interactive response for each prompt".¹⁹¹⁹ Google said its implementation "can sometimes take a minute or more to generate results". It also reported "occasional inaccuracies in the outputs".¹⁹¹⁹

Google tested the results with human raters. Sites designed by human experts "had the highest preference rates". Google's generative UI followed closely, "with a substantial gap from all other output methods".¹⁹¹⁹ Google noted that this evaluation "did not take into account generation speed".¹⁹¹⁹

About 28 months separate Anthropic's Artifacts from Intelligent UI. About 11 months separate it from Google's generative UI.²¹²¹21Arithmetic in calendar months: June 2024 to October 2026 is 28 months, and November 2025 to October 2026 is 11 months. https://www.anthropic.com/news/claude-3-5-sonnet and https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/

The API offers the same model names without the interface. OpenAI's API pages price GPT-6 Sol at $2.00 per million input tokens and $10.00 per million output tokens. GPT-6 Luna costs $0.10 and $0.50.²²²²22OpenAI API docs, model pages for GPT-6 Sol and GPT-6 Luna, undated, read on 17 October 2026: Sol at $2.00 input, $0.20 cached input and $10.00 output per 1M text tokens; Luna at $0.10, $0.01 and $0.50; a 1,050,000-token context window and 128,000 max output tokens for both; input "Text, Image" and output "Text"; knowledge cutoff Apr 20, 2026 for Sol and May 18, 2026 for Luna; Sol "Built to power complex coding and agentic workflows", Luna "Our most efficient model for focused, high-volume tasks". https://developers.openai.com/api/docs/models/gpt-6-sol and https://developers.openai.com/api/docs/models/gpt-6-luna On both input and output, Sol costs 20 times as much as Luna.²³²³23Arithmetic: $2.00 divided by $0.10 is 20, and $10.00 divided by $0.50 is 20. https://developers.openai.com/api/docs/models/gpt-6-sol and https://developers.openai.com/api/docs/models/gpt-6-luna Both models have a context window of 1,050,000 tokens and can return up to 128,000 tokens.²²²² OpenAI pitches Sol as "Built to power complex coding and agentic workflows" and Luna as its most efficient model for focused, high-volume tasks. The knowledge cutoff is 20 April 2026 for Sol and 18 May 2026 for Luna.²²²²

Both API pages list text and images as input, and text alone as output. They list no output type for an interface.²²²² None of the OpenAI sources read offers Intelligent UI in the API or in the Apps SDK.⁶⁶

No catalogue of the components GPT-6 can use in ChatGPT appears on OpenAI's pages.⁶⁶ It has not said whether interactive components run in a sandbox. It has not said whether they can open links or reach the network.⁶⁶ No statement covers data typed into a form inside an answer, or code from third parties.⁶⁶ No OpenAI source gives the wait before an answer in seconds or milliseconds.⁶⁶ No OpenAI page says that Intelligent UI costs extra or has a usage limit of its own.⁶⁶