ModelsArchitectures & capability
ChatGPT's new Intelligent UI did not trigger on OpenAI's showcase prompts in a hands-on test
OpenAI's GPT-6 lets ChatGPT mix text with charts, forms and small tools, deciding on its own when to do so. One hands-on tester found OpenAI's showcase prompts returned plain text until the model was explicitly told to use its visualization skill.
RisingHow it’s measured
YouTube video · 7.2K views · 49 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisOpenAI launched GPT-6 with Intelligent UI on October 7, 2026. The feature lets ChatGPT build answers that combine text with interactive elements such as charts, calculators, maps and forms, and the model decides when to use them. Paid tiers get GPT-6 Sol, and Free and Go users got GPT-6 Luna a day later. GPT-6 Astra, which powers the Pro reasoning level, does not support the feature, and neither do older desktop apps. [4] [5]
In one informal test, the bike-design and road-trip prompts OpenAI uses as examples came back as ordinary text twice. Interactive visuals appeared only after the tester told ChatGPT to use its visualization skill. The same tester said Claude made its visual on the first try, but judged ChatGPT's result more detailed and interactive once it appeared. [1] [3]
The vendors build these interfaces in different ways. OpenAI assembles answers from ready-made components that can stream, with a compiler drawing the interface while the model is still generating. Claude writes HTML/SVG directly. Google's generative UI relies on tools, detailed system instructions and post-processing. No public benchmark yet compares how often each one activates or how good the result is. [5] [6] [7]
the feature is live from Free through Enterprise tiers, and OpenAI admits its design judgment still needs work.
Read the full assessment
Implication: teams that need consistent interactive output should ask for the format explicitly and check app versions, admin settings and Pro-level exclusions.
Executive brief
The biggest problem with ChatGPT's new "Intelligent UI" is getting it to appear. A Reddit commentary post and its linked video (published October 9, 2026) say the feature did not switch on for OpenAI's own showcase prompt. The tester says it failed twice and only produced interactive visuals after being told to use its "visualization skill" (video). OpenAI launched GPT-6 with Intelligent UI on October 7. The feature mixes text with charts, buttons, forms and small tools (Search Engine Journal). OpenAI says the model decides when to use it and admits its "design judgment" still needs work (TestingCatalog). Only one person's informal test backs the triggering claim.
What changed and event timeline
Google ships generative UI
Gemini 3 arrived with two experiments, "dynamic view" and "visual layout," which build an interactive interface for each prompt. Google Research published an evaluation the same day ().
Claude adds charts and diagrams inside replies
Claude draws HTML/SVG visuals inside its replies, either on its own judgment or when asked ().
GPT-6 and Intelligent UI launch
Plus, Pro, Business and Enterprise users get GPT-6 Sol. OpenAI published a system card the same day (;).
Hands-on commentary reports triggering failures
The bike and road-trip prompts came back as plain text until the tester asked explicitly (;).
Capabilities and access
- Models: GPT-6 Sol on paid tiers and GPT-6 Luna on Free and Go.
- Where it works: reasoning levels from Instant to Extra High, on the web and updated apps. Enterprise access depends on admin settings (Search Engine Journal).
- Outputs: cooking timelines, road-trip maps, savings calculators, bill splitters and games (Tech Insider).
Read the full section
- Models: GPT-6 Sol on paid tiers and GPT-6 Luna on Free and Go. GPT-6 Astra powers the "Pro" reasoning level and does not support Intelligent UI (Search Engine Journal).
- Where it works: reasoning levels from Instant to Extra High, on the web and updated apps. It does not work on older macOS and Windows desktop apps. Enterprise access depends on admin settings (Search Engine Journal).
- Outputs: cooking timelines, road-trip maps, savings calculators, bill splitters and games (Tech Insider).
- Settings: a "Layout and visuals" setting on the web turns it off, though some elements may still appear (Implicator).
Technical analysis for researchers and developers
- Design difference: OpenAI's version assembles answers from ready-made, streamable components. Google's system combines tools, detailed system instructions and post-processing (Google Research).
- Ready-made components: they should load faster and look more consistent, but they limit what the model can build.
- Format choice: the model picks the format for each prompt, so a prompt with no interactive output may be the model choosing text, not a bug (MindStudio).
Read the full section
- Design difference: OpenAI's version assembles answers from ready-made, streamable components. A compiler renders the interface while the model is still generating, and GPT-6 can interleave reasoning with partial output (TestingCatalog). Claude instead writes HTML/SVG directly (The Decoder). Google's system combines tools, detailed system instructions and post-processing (Google Research).
- Ready-made components: they should load faster and look more consistent, but they limit what the model can build.
- Format choice: the model picks the format for each prompt, so a prompt with no interactive output may be the model choosing text, not a bug (MindStudio).
- Missing evaluation: no published benchmark measures how often the feature switches on, or how good the result is.
Claims and evidence
- Training claim: OpenAI says GPT-6 was trained to decide when interactive formats help. This is OpenAI's statement and has not been independently tested (Search Engine Journal).
- Speed claims (OpenAI's figures): GPT-6 Instant starts answering web-search questions 44% sooner than GPT-5.6 Instant, and GPT-6 Extra High is as fast as GPT-5.6 Medium (TestingCatalog).
- Triggering failures: these come from one tester, with no prompts or logs published (video). Other independent reports of failures could not be found.
Read the full section
- Training claim: OpenAI says GPT-6 was trained to decide when interactive formats help. This is OpenAI's statement and has not been independently tested (Search Engine Journal).
- Speed claims (OpenAI's figures): GPT-6 Instant starts answering web-search questions 44% sooner than GPT-5.6 Instant, and GPT-6 Extra High is as fast as GPT-5.6 Medium (TestingCatalog).
- Triggering failures: these come from one tester, with no prompts or logs published (video). Other independent reports of failures could not be found.
- Claude vs ChatGPT: the tester says Claude produced its visual on the first try, but ChatGPT's result looked and worked better once it appeared (video). These are personal impressions.
Context and prior work
Google treated generative UI as a research problem. Its PAGEN dataset is a set of websites designed by experts. Raters preferred generative-UI outputs over standard text and web results, ranking them second only to the expert designs (Google Research). The video says Claude added its visuals "last year" (video). Anthropic actually launched them in March 2026 (The New Stack).
Limitations, safety and contested findings
- OpenAI's own caveat: design judgment and the range of possible interfaces "still need work" (TestingCatalog).
- Safety regressions in the system card: GPT-6 Sol scored worse on adult self-harm tests.
- Errors in the commentary: the video calls the previous model "GPT-5.5," but reporting names GPT-5.6 (video).
Read the full section
- OpenAI's own caveat: design judgment and the range of possible interfaces "still need work" (TestingCatalog).
- Safety regressions in the system card: GPT-6 Sol scored worse on adult self-harm tests. In teen tests, emotional-reliance scores fell from 0.927 to 0.734 for Luna and from 0.921 to 0.770 for Sol. OpenAI says an extra teen classifier, not included in these results, covers part of the gap (Implicator).
- Errors in the commentary: the video calls the previous model "GPT-5.5," but reporting names GPT-5.6 (video). The post also sells a "GPT-6 Astra" course, yet Astra does not support Intelligent UI (video).
Business and practitioner implications
- Prompt explicitly: if consistent output matters, ask for the format, for example by telling ChatGPT to use its visualization skill (video).
- Publishers: calculator and converter websites face disruption as users build these tools inside the chat (Search Engine Journal).
- Enterprise teams: check admin settings and app versions before rollout. Teams that rely on Pro (Astra) reasoning won't get the feature.
Read the full section
- Prompt explicitly: if consistent output matters, ask for the format, for example by telling ChatGPT to use its visualization skill (video).
- Publishers: calculator and converter websites face disruption as users build these tools inside the chat (Search Engine Journal).
- Enterprise teams: check admin settings and app versions before rollout. Teams that rely on Pro (Astra) reasoning won't get the feature.
The source trail.
Sources (11)
I Tested ChatGPT Intelligent UI Twice: Here's the Real Catch
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=vC7rYzjDjAU); automatic text may contain errors.
reddit.comGPT-6 Astra Intelligent UI is INSANE!
Related coverage; assess separately
reddit.comLinked video: I Tested ChatGPT Intelligent UI Twice: Here's the Real Catch
linked video auto captions
www.youtube.com