On 7 October OpenAI published a launch page with a quiet little experiment on it. The same request goes to two models: “Give me a plan for a sunday lamb roast, I’m having friends over still figuring out numbers tbh.”
The older model answers the way we are used to. Headings, a table of weights per guest count, a timeline in bullet points. The new one answers with a product. Photos of the dishes. A stepper labelled “Number of people” with a minus and a plus. A shopping list that recalculates when you tap. An eight-step cooking checklist that counts “1 of 8”. A button that says “Copy shopping list”.
Nobody at OpenAI designed that screen for that roast. OpenAI designed the parts, and the model decided how they fit together for one person, once. I think that small demo is the most consequential design announcement of the year, and almost nobody in design is talking about what it removes.
The Ledger
Figma’s agent left beta and got a meter. On 6 October Figma moved the agent in Figma Design and its Weave tools from open beta to general availability, and from that day they consume AI credits from the same budget as Figma Make. To soften it, Figma says AI credit add-ons now carry 2x more credits on Professional and 1.6x more on Organization and Enterprise at the same price, and admins can now set a custom credit limit per user.
Why it matters for practitioners: agentic design work now has a unit cost that finance can see per person. Expect the question “who on the team gets more credits” within a quarter, and decide your answer before someone outside design does.
OpenAI will watermark ChatGPT text in Europe, and admits editing weakens it. To comply with the EU AI Act’s transparency rules, OpenAI says it will roll out an invisible text watermark to ChatGPT and Codex users in the EU over the coming weeks, off by default in the API. In its own tests, swapping 10% of the words for synonyms dropped detection from about 92% to 66%, and OpenAI warns that a missing watermark “does not prove human authorship.”
Why it matters for practitioners: provenance is becoming a property of the words themselves, and it travels through copy and paste. If your product lets people draft with AI, your editing, export and sharing flows are now part of a compliance story whether you planned it or not.
Ads are moving into the moment of creation. OpenAI announced a visual ad format that it will test during image generation in ChatGPT for Free and Go users, starting in the US later this month, labelled and kept separate from the generated image. In the same post it cites an early partner result: WeightWatchers’ attributed cost per acquisition on ChatGPT Ads was 15.3% lower than its blended paid-search benchmark, according to DV Rockerbox.
Why it matters for practitioners: when the response is a composed interface rather than a page, the ad slot is also a design decision made at runtime. Brand teams should ask now where their products will appear next to generated answers, and who decides the layout.
Google shipped a meeting note-taker that never leaves the laptop. Google AI Edge Foresight is a Mac app that transcribes meetings, writes notes and answers questions entirely offline, using the on-device EmbeddingGemma 2 model with 740 million parameters and a Gemma 4 assistant. TechCrunch frames it as a Granola competitor in a crowded category.
Why it matters for practitioners: “it runs on your machine” is turning into a product feature you can design around, not just an engineering choice. For regulated clients, local-first may now be the shortest path past the security review.
The Interface Nobody Sees First
Read how OpenAI describes what it built. It made “a library of native, streamable components, along with a compiler that processes the interface as the model generates it.” The library gives each response “a familiar design foundation while leaving the model room to decide how the pieces fit together.” Then the model was trained to “make good design decisions, including how to organize information clearly, when to use interactivity, and when a simple text response is enough,” and the interfaces it created were evaluated “for clarity, usefulness, and completeness.”
And then, honestly: “There’s still work ahead to improve the model’s design judgment.”
The page ends with a line every product leader will hear quoted back at them by Christmas: “For decades, people have had to learn how to use software, navigating fixed interfaces designed around particular tasks. We believe the future will be different. Instead of people adapting to software, software will adapt to people.”
Here is the contradiction I want to sit with. OpenAI promises software that adapts to people, and the price of that promise is that no designer sees the interface before the person does.
What review was built for
Almost every quality practice in design assumes an artifact you can look at before it ships. The crit. The usability test. The accessibility audit. Legal reading the copy. Brand signing off the colours. Even an A/B test, which already broke the idea of one screen, still has a finite set of variants someone approved.
A composed interface is not an artifact. It is a distribution of artifacts. The lamb roast with a guest stepper is one sample from it. The next person asking about a roast for twelve with two vegetarians gets a different screen, and nobody at the company that shipped it has seen that one either. You cannot review a distribution by looking at a screen. You can only review the things that produce it.
That is the part I would want a head of design to absorb this month. GPT-6 in ChatGPT is not a niche experiment. OpenAI says it is rolling out to more than 1.2 billion weekly users, paid tiers first and Free and Go from the following day. Your customers will spend the next year getting used to answers that arrive as small, disposable products. Their patience for the fixed form you shipped in March will change accordingly.
To be fair to the other side
The case for this is strong, and I do not want to pretend otherwise.
Fixed interfaces never served everyone. They served the average user the team imagined, and they failed people at the edges: the screen reader user, the person with low literacy, the one on a small phone in a second language. A composed interface can, in principle, meet each of them where they are. And “nobody saw it first” was already partly true. Personalisation, feature flags and recommendation slots meant no one at a large company had seen every version of their own product for years.
OpenAI also did not remove design. It moved it. Someone designed that stepper, that checklist row, that photo card, and someone wrote the criteria the model was graded against. I think that is exactly right, and it is the point. The design did not disappear. It went upstream, into parts, rules and tests, and most design organisations are not staffed or measured for any of the three.
Where the craft goes
The same week made the destination unusually concrete. When Figma’s agent went generally available, it came with a feature that tells you what a design system is about to become. Library owners can now attach guidelines files that the agent reads in the background whenever someone asks it to design with that library. Figma’s own help page recommends files like overview.md, components.md and anti-patterns.md, three to eight of them, written as scenarios: what to do, what not to do, and why. The combined size of all guidelines in a library cannot exceed 200 KB. And because they add context, Figma notes that a prompt using a library with guidelines may consume more AI credits.
Sit with those three facts together. Your design system is becoming a text file a machine reads every time. It has a size limit. And every rule in it has a per-prompt price.
Ethan Mollick gave the best small example of why this matters in his September essay on what he calls the capability overhang. AI, he writes, has “a very annoying habit of adding little headlines on top of your headlines,” and if you have no grounding in design “you may not realize that you need to ask the AI to stop ‘adding eyebrows’ to the work.” That is a design system rule in miniature. A model’s taste has defaults. Someone with vocabulary has to name them and write them down, or the defaults become your brand.
So I see three layers where design leadership now lives. The parts: the components a model is allowed to use, built to look right in any arrangement, not just the ones in the mockup. The grammar: the rules for composing them, written as decisions with reasons, short enough to fit and cheap enough to run. And the exam: the set of real requests and the criteria that decide whether a generated answer is good enough to show a customer. OpenAI’s three words, clarity, usefulness and completeness, are a start. Your product needs its own, and someone in design has to own them the way a QA lead owns a test suite.
This changes what founders and CEOs should ask their design leaders for. Not more screens. A component library that survives recombination, a grammar a machine can follow, and an evaluation the team runs every week and reports like any other quality metric. If you ship anything inside other people’s assistants, or plan to let a model assemble your own UI, those three things are your brand. The screen is only where it shows up.
Monday action
Write the exam before you write the next screen. Pick one answer your product gives often enough to matter: a pricing explanation, an onboarding step, a support reply, a report summary. Collect 20 real requests that lead to it. For each, write two lines: what a good response must contain, and what it must never do. Then let a model compose responses from your existing components, through the Figma agent, your own stack or any assistant, and have one designer grade all 20 against clarity, usefulness, completeness, and one criterion of your own: would we have shipped this screen? The pattern in the failures is your first guidelines file. Keep it under a page.
I would like to know which rule you end up writing first, because I suspect it will say more about your product than your last brand guidelines did. Reply and tell me, or tell me why you think the fixed screen still wins in your market.
Sources
OpenAI, “GPT-6 and Intelligent UI for everyone,” 7 October 2026: https://openai.com/index/gpt-6-for-everyone/
Figma, “Add guidelines files to a design library,” Figma Learn help center (accessed 13 October 2026): https://help.figma.com/hc/en-us/articles/43509372175511-Add-guidelines-files-to-a-design-library
Figma, “AI credit updates FAQ,” Figma Learn help center (agent general availability on 6 October 2026): https://help.figma.com/hc/en-us/articles/42614902212887-AI-credit-updates-FAQ
Ethan Mollick, “The Overhang,” One Useful Thing, 18 September 2026: https://www.oneusefulthing.org/p/the-overhang
TechCrunch, “OpenAI will start watermarking ChatGPT’s text in the EU,” 5 October 2026: https://techcrunch.com/2026/10/05/openai-will-start-watermarking-chatgpts-text-in-the-eu/
OpenAI, “Building advertising for the way people use AI,” 5 October 2026: https://openai.com/index/new-chatgpt-ads-format-and-measurement/
TechCrunch, “Google releases a new local-first Granola competitor,” 8 October 2026: https://techcrunch.com/2026/10/08/google-releases-a-new-local-first-granola-competitor/
Germán