Shuichiro Ogawa
日本語

Notes · updated 2026-09-28

AI and Design Weekly Watch (2026-09-21 to 09-28)

An integrated summary of 'AI and design' developments over the past seven days, collected in three tiers: T1v vendor primary sources, T2 public institutions and standards, and T3 expert opinion.

Contents (5)
  1. T1v Vendor Primary Sources
  2. T2 Public Institutions and Research
  3. T3 Personal Views of Credible Individuals
  4. Recent Major Updates (Chronological)
  5. How to Read the Reliability Tiers

Developments in “AI and design” over the past seven days (2026-09-21 to 2026-09-28) were collected in three tiers by source reliability. The gap since the previous watch (AI and Design Weekly Watch (2026-09-14 to 09-21)) is exactly seven days, with no missed week. The scope covers four areas: integration of generative AI and design tools, AI adoption in design practice, design-related announcements from major AI vendors, and research and regulatory developments. Items already known from the previous watch are not repeated in this note. Starting with this issue, the full text of every reference was retrieved, and figures and dates were checked against that text before writing.

Related: AI and Design Weekly Watch (2026-09-14 to 09-21) (the previous weekly watch) / Design Agent Tools in 2026: The Current State of Autonomous Production (the current state of autonomous production agents) / MCP and Design Systems: The Infrastructure Layer for Agent Integration (connecting design tools via MCP) / AX and Design: A Cross-Comparison of Academic and Industry Perspectives (contrasting academic and industry views of AX).

Last week, the vendor side leaned toward distribution and education. This week, design tools shipped ways for agents to learn how a team works. Framer released Skills, which save and reuse instructions for a design system or writing style, and in the same week Figma’s VP of Product wrote about turning repeated work into skills and plugins the team can reuse. Webflow let agents build interactions with the same triggers as the canvas. On the model supply side, Anthropic’s Claude Opus 5.5 shipped and reached Lovable the same day. OpenAI shut down the Sora 2 API as announced and named no replacement.

T1v Vendor Primary Sources

Framer released Skills, a way to teach the Framer agent how you work (09-22). Instructions for a design system, writing style, or CMS workflow can be saved and reused across tasks. Users ask the agent to create or update a skill with /skills, and can have it reference @pages, @components, and @styles in the project. The official update page suggests asking the agent, after a conversation that refined a page or copy, to capture the decisions worth reusing next time. Skills stay with a project when it is remixed, so template creators can bundle guidance for adding pages without breaking the design.

At Figma, VP of Product Paige Costello discussed the work that remains once a prompt can become something clickable in a few hours (09-24). It is an opinion piece by a product executive, not a feature announcement. The article splits a team’s repeated work into two forms. Skill: a reusable set of instructions triggered with a / command (for example, checking a design against accessibility guidelines). Plugin: a reusable tool that performs a specific task on the canvas (for example, generating a dashboard layout). It also suggests prompting the Figma agent to build the plugin itself. Framer’s feature and Figma’s article take different forms, but both encourage saving a team’s judgment in a form agents can read and reusing it on the next job.

Webflow let agents create and edit interactions on a page through MCP (09-21). Interactions are built on IX3 and GSAP and use the same scroll, hover, load, and click triggers as the canvas. The same update lets MCP sort and filter CMS content in a single request, deploy and debug Webflow Cloud apps, and pull and merge branches (Enterprise only).

Midjourney updated its alpha site (changelog dated 09-23, published 09-24). Turning on “Live previews” in the Styles sidebar makes fast models preview the current prompt in each style. “Set current pills as default” saves default parameters, and the Create Feed became a full-width masonry layout. The editor now supports v8.1 and v8.2 edit types, and edit canvases can run in HD. Korean, which came to alpha last week, is now live for everyone on midjourney.com.

On model supply, Anthropic released Claude Opus 5.5 (09-22). Pricing is $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5. Cache reads cost $0.20, 60% less. The announcement’s comparison, that it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5,” is the company’s own assessment and has not been independently verified. Claude Code made Opus 5.5 the default Opus model in v2.1.280 (09-22). Lovable added Opus 5.5 the same day and reported that, on its own benchmarks against Opus 5, steps per task fell by 26 to 48% for building from zero and by 34 to 47% for iterative code fixing. Lovable writes that every difference in its tables is statistically significant at the 95% level, but it has not published the benchmark tasks or the number of runs.

Lovable went on to let the AI features built into apps use Claude models (Fable 5.1, Opus 5.5, and others) without an Anthropic account (09-23), and added OpenAI’s GPT-6 Sol and GPT-6 Luna (09-25). GPT-6 Sol and Luna also arrived on Vercel’s AI Gateway (09-22). OpenAI’s own GPT-6 announcement was not retrieved this time, so availability is confirmed only as far as the Vercel and Lovable entries go. On 09-25, Vercel released “Pixel Canary,” a model whose provider is not disclosed, free on AI Gateway. The claim that it is strong at coding, frontend development, and mobile app design comes from the company’s description, with no evaluation method given.

Lovable added a daily free chat allowance to Free, Pro, and Business workspaces (09-24). Generating images or video in chat, or handing work off to Plan or Build, uses credits, and this pricing applies through 2026-10-31. On 09-25, web sources cited in an answer began appearing as small pills with the site’s icon and domain. Hovering shows the page title, an excerpt, and the URL.

OpenAI removed the Videos API and the Sora 2 models (sora-2, sora-2-pro, and dated snapshots) from its API on 09-24, as announced on 2026-03-24. In the deprecations table, the recommended replacement column is blank. Within OpenAI’s API, no replacement for Sora 2 is indicated.

Some vendors could not be confirmed for the window. According to the collecting subagent, Figma’s release notes and Canva’s newsroom had no new items in the window. The dates of Adobe’s September Firefly updates could not be pinned down, and the official pages of Stability AI, Replit, xAI, Google DeepMind, and Microsoft could not be reached. These are “undetermined,” not “no announcement.”

T2 Public Institutions and Research

On the regulatory and standards side, no new publication, revision, or stage advance in the window could be confirmed from primary sources. The EU Code of Practice on transparency still shows about 190 signatories as of the end of July 2026, with the page last updated on 07-31. The Article 50 guidelines were last updated on 08-06, and Article 50 itself has applied since 08-02. The UK IPO designs consultation still reads “We are analysing your feedback,” and Part 3 of the U.S. Copyright Office’s “Copyright and AI” remains the pre-publication version of 2025-05-09. The Study Group on Japan’s AI Guidelines for Business still lists its 10th meeting (by email, 08-07 to 08-13, materials published 08-27) as the latest, and the latest C2PA specification is still 2.4. NIST CSRC news in the window covered a draft OT security guide (09-21) and a threshold cryptography event (09-25), neither related to AI and design.

Last week, the report that Phase 2 of the NIST GenAI Text Challenge would start on 09-28 was held over as a candidate to check. Fetching NIST’s challenge page this time showed both the Phase 2 start and close as “TBD.” The 09-28 start appears only in search-result summaries and is not supported by the primary page.

On the Agency for Cultural Affairs page “AI and Copyright,” the latest date in the retrieved text is 2024-08-09 (Reiwa 6), with no additions in the window. However, the previous note recorded a document dated 2025-09-11 (Reiwa 7) as the latest, which conflicts with this retrieval. Whether this is a gap in retrieval or an error in the earlier record has not yet been resolved.

Some items remain on hold because primary sources could not be reached. The official pages for USPTO Federal Register Notices, the W3C AI Content Disclosure Community Group, ISO/CD 22144, the Japan Patent Office’s design system subcommittee, and EUIPO’s generative AI guideline update could not be retrieved.

No analyst research was adopted this time either. No design-related research report with a stated methodology was published in the window. The nearest candidates all fell outside the window, and many were published by sellers of or investors in AI tools. The list of candidates considered and excluded is in the corpus.

T3 Personal Views of Credible Individuals

In his 09-25 UX Roundup, Jakob Nielsen covered a study by Kodandaram and colleagues at Stony Brook University and Old Dominion University. The researchers wrapped a computer-use agent in a screen-reader-accessible shell (OLLA), and eight blind users ran it on their own PCs for three weeks. In Nielsen’s summary, the GPT-5-based agent fully completed 53% of 1,258 commands issued across 12 applications. Nielsen sets this against screen-reader users’ task completion rates of roughly 25 to 55% in other studies and judges the agent already competitive. Citing the best OSWorld score’s growth of 33 points a year, he predicts that even at a third of that pace, 10 points a year, agents for blind users will pass the 72% sighted baseline in 2028. The comparison places separate studies side by side; they were not measured under the same conditions. All figures come from Nielsen’s summary, and the original papers were not checked.

The same issue drew on two UIST ‘26 papers to present synthetic users as a tool for designer reflection. In SiMUSation, by Huanchen Wang and colleagues at Southern University of Science and Technology, a museum designer lays out exhibits on a 2D floor plan and releases LLM-driven visitor personas into it. In Nielsen’s summary, the personas’ behavioral plausibility scored only 4.8 out of 7, while design insight scored 5.6. Four expert judges who did not know which layouts were revisions rated the revised layouts about two-thirds of a point higher. The other paper, AER, decomposes a reference image into technical and conceptual tags and lets artists generate variations of their own work from the tags they choose. In a within-subjects comparison with 16 professional artists, agency was 0.9 points higher on a 7-point scale than with a Midjourney-style baseline (Cohen’s d = 0.85). In a two-week field study, however, four artists spent 9.4% of their time on the feature that asks simulated critics for feedback. Nielsen reads this as showing that simulated reactions cannot replace observing real people, but can prompt designers to reconsider their choices.

The 09-21 issue covered a preprint by Jorge Fábrega of Universidad del Desarrollo analyzing the Anthropic Economic Index. It divides ways of directing AI work into “specified delegation,” where instructions, constraints, and acceptance criteria are set before execution, and “iterative coproduction,” where the user steers by correcting the AI’s drafts. In Nielsen’s summary, when 10 points of usage shift from personal to work purposes, the degree of specification (0 to 100) rises by 2.8 points in API traffic and 1.5 points in Claude.ai. Nielsen reads this as users writing briefs up front when the outcome carries professional consequences. These are preprint figures, via Nielsen’s summary.

Luke Wroblewski argued for reversing the order of building a tool and then getting the outcome, and for using AI to deliver the outcome directly (09-24). He writes, “A tool was always just a means to an end. If AI can get you to the end directly, why stop at the tool?” His evidence is Exposit, an AI newsroom he worked on. It was built as a news site, but once a feature was added that compiles a personalized report on a requested topic, that quickly became the dominant way people used the site. No figures such as usage shares are given. Where Framer and Figma point toward giving agents judgment inside the tool, Wroblewski points toward skipping the tool altogether.

In a short note, Simon Willison wrote that the more time he spends with coding agents, the more convinced he is that they make software engineering even harder (09-24). Unlocking their full potential, he says, requires “extraordinary discipline and knowledge.” The note does not address design directly. Still, it overlaps with Costello’s point that getting there fast does not mean you have built something great. Neither cites the other, and both are opinions.

Recent Major Updates (Chronological)

  • 2026-09-25: Lovable adds GPT-6 Sol/Luna and source pills; Vercel releases Pixel Canary for free (T1v) / Nielsen’s UX Roundup 09-25 issue (T3)
  • 2026-09-24: OpenAI removes the Videos API and Sora 2 from its API; Lovable launches a free chat allowance; Midjourney publishes its alpha changelog; Figma’s Costello publishes her opinion piece (T1v) / Wroblewski’s “Skip the Tools,” Willison’s note (T3)
  • 2026-09-23: Lovable adds Claude models to app AI features (T1v)
  • 2026-09-22: Anthropic releases Claude Opus 5.5; Claude Code v2.1.280; Lovable adds Opus 5.5; Framer releases Skills; GPT-6 Sol/Luna arrive on Vercel AI Gateway (T1v)
  • 2026-09-21: Webflow supports creating interactions via MCP (T1v) / Nielsen’s UX Roundup 09-21 issue (T3)

How to Read the Reliability Tiers

  • T1v (vendor primary): Full text was retrieved and checked for Framer, Figma, Webflow, Midjourney, Anthropic, Lovable, OpenAI, and Vercel. Webflow could not be fetched directly, and its text was checked through a reader proxy. Three items were treated as self-assessment: Anthropic’s comparison with Fable 5.1, Lovable’s in-house benchmarks, and Vercel’s description of Pixel Canary. Figma’s article is a product executive’s view, not a feature specification. OpenAI’s primary source for GPT-6 was not retrieved.
  • T2 (public institutions, standards, research): None of the nine retrieved items changed in the window. For NIST, the primary page says “TBD,” correcting the information carried over from last week. The Agency for Cultural Affairs discrepancy with the earlier record remains. No research was adopted.
  • T3 (expert opinion): Four pieces by three people. All study figures cited in Nielsen’s two issues come via his summaries, and the original papers were not checked.

Scope and Method of the Survey

Ledger details (position judgments, exclusion records, the list of unreachable sources, and per-claim verification evidence) are in the corpus (source/review/ai-design-watch-2026-09-28/ai-frontier.md). The retrieval record for the full text of all 24 references is in fulltext-manifest.json in the same directory.

Unverified Items

  • Adobe’s September Firefly updates: the official page could not be retrieved, so whether they fall in the window is undetermined.
  • Stability AI, Replit, xAI, Google DeepMind, and Microsoft Designer: official pages unreachable, so whether they had new items in the window is undetermined.
  • OpenAI’s own announcement of GPT-6 Sol/Luna was not retrieved; availability is confirmed only via Vercel and Lovable.
  • The latest document date found on the Agency for Cultural Affairs page (2024-08-09) does not match the earlier record (2025-09-11).
  • The start date of NIST GenAI Text Challenge Phase 2 is “TBD” on the primary page.
  • USPTO, the W3C AI Content Disclosure CG, ISO/CD 22144, the JPO design system subcommittee, the Agency for Cultural Affairs’ 26th-term policy subcommittee, and EUIPO: primary sources not reached.
  • The publication date of McKinsey’s “The State of AI: Global Survey 2026” (08-25) rests only on a subagent report; the text was not retrieved.
  • The original papers behind the studies Nielsen summarized (OLLA, SiMUSation, AER, and Fábrega’s preprint) were not retrieved.
  • Lovable’s benchmark tasks and number of runs are not public.

References

All accessed 2026-09-28.

T1v Vendor Primary

T2 Public Institutions and Standards

T3 Expert Opinions

Author: Shuichiro Ogawa (Design Researcher / Consultant) About me →