2026-06-23

Claude in 2026: New Features, the Design Shake-Up, and What It Means for PMs

Claude Code, Cowork, and a dedicated Claude Design tool have shipped fast — and they're reshaping how design work gets done. Here's what's real, what's changing for designers, and what it means for PMs who now prototype themselves.

TL;DR: Claude has shipped fast in the last year — Claude Code, Cowork, Managed Agents, Claude for Excel/PowerPoint, and a dedicated visual-creation tool called Claude Design — and a chunk of that is specifically reshaping how design work gets done. Designers have gone from cautious AI users to power users (91% weekly, up from 54% a year ago), the UX job market has genuinely polarized (job postings down 71% from their 2022 peak, but AI-fluent senior roles growing), and a real "PM-as-builder" trend has emerged where product managers prototype directly in Claude instead of waiting on a designer. None of this replaces design judgment — it relocates where that judgment gets applied.

Executive Summary

What Claude has shipped recently: Claude Code (now a $2.5B+ run-rate business), Claude Cowork (agentic desktop work), Claude Managed Agents, Claude for Excel/PowerPoint, and model upgrades through Opus 4.8 and the new "Mythos-class" Fable 5 — plus MCP being donated to the Linux Foundation as shared industry infrastructure.
Claude's design-specific tooling: Artifacts (interactive prototypes/dashboards), a dedicated product called Claude Design (1M+ users in its first week), and official integrations with Figma and Adobe Creative Cloud. Important correction: Claude does not natively generate raster images like DALL-E or Midjourney — its visual strength is code-based (SVG, interactive prototypes, diagrams), not pixel generation.
What's happening to the design profession: real polarization, not uniform displacement. UX job postings are down 71% from their 2022 peak and one in three orgs cut UX staff in 2024 — but AI-fluent, senior design hiring at AI-native companies actually rose ~60% in 2025. The durable core of the job — taste, judgment, stakeholder navigation — is becoming more valuable, not less, as AI lowers the bar for producing interfaces.
What's happening to PMs specifically: a genuine "PM-as-builder" trend, with named examples at Meta and LinkedIn (which replaced its APM program with an "Associate Product Builder" track requiring a built product). The skill gap that matters most for PMs is reviewing AI-built artifacts against intent and risk, not learning to code.

Background / Context

Two things are true at once in 2026. First, Anthropic has been shipping at a pace that makes "what's new with Claude" a moving target — model releases roughly every 6-8 weeks, plus a wave of office/creative-tool integrations. Second, a meaningfully overlapping set of those launches is aimed directly at visual/design work: Artifacts, Claude Design, and official Figma/Adobe connectors. That overlap is why "new Claude features" and "AI changing design roles" are really one story, not two.

Key Findings

What Claude has actually shipped (mid-2025 - June 2026)

Claude Code launched publicly in May 2025 and became Anthropic's fastest-growing product — it crossed $1B annualized revenue by November 2025 and reportedly grew past $2.5B by mid-2026, with weekly active users doubling since January 2026.
Claude Cowork (research preview Jan 2026, GA on macOS/Windows April 9, 2026) brings agentic, Claude-Code-style task execution to general desktop knowledge work rather than just coding — it can access local files/apps but asks permission per application and shows its work step by step.
Claude Managed Agents (launched April 8, 2026) is a hosted API where Anthropic runs the agent harness/sandbox/session logs for you, priced per session-hour plus tokens; Notion, Rakuten, Asana, Sentry, and Atlassian were production customers at launch.
Claude for Excel (GA to all Pro subscribers Jan 24, 2026) and Claude for PowerPoint (preview Feb 2026, sharing context with Excel since March 2026) put Claude directly into Office workflows — multi-tab workbook queries, cell-level citations, model-building from templates.
Claude.ai Skills (launched Oct 2025, "Skills 2.0" in Q1 2026) package reusable workflows — including prebuilt Skills for Excel/PowerPoint/Word/PDF — that Claude loads only when relevant.
MCP (Model Context Protocol) was donated to the Linux Foundation in December 2025, with OpenAI, Google, and Microsoft as co-sponsors of the new "Agentic AI Foundation" — effectively ending "MCP as an Anthropic thing" and making it shared infrastructure, with 17,000+ active public MCP servers counted by Q1 2026.
Model releases: Opus 4.6 (Feb 2026, 1M context), Opus 4.7 (April 2026, major coding/vision gains), Opus 4.8 (May 2026, currently leads coding benchmarks like SWE-bench Verified at 88.6%), Sonnet 4.6, Haiku 4.5, and Claude Fable 5 (announced June 9, 2026) — Anthropic's first "Mythos-class" tier, priced at $10/$50 per million tokens.
Platform reach: Claude is now available on all three major clouds (AWS, Google Cloud, and — via a November 2025 deal — Microsoft Azure/Foundry), and Anthropic's overall annualized revenue reportedly crossed $47 billion by May 2026, alongside a confidential IPO filing.
Competitive snapshot: as of June 2026, Opus 4.8 leads the Artificial Analysis Intelligence Index and coding benchmarks; GPT-5.5 leads on agentic-workflow benchmarks and creative writing; Gemini 3.1 Pro leads on math/reasoning and price-to-performance.

Claude's design-specific tooling

Artifacts evolved from a side-panel code/document preview (2024) into a fuller environment: GA across all plans and mobile (Nov 2025), "Live Artifacts" with auto-refreshing data connections (April 2026), and "Claude Code Artifacts" (June 2026) that turn a coding session into a shareable, self-contained HTML page — explicitly described by Anthropic as "a capture of work, not an application."
No native image generation. This is worth stating plainly because it's a common assumption: as of June 2026, Claude cannot generate raster images (JPG/PNG) the way DALL-E, Midjourney, or Gemini's image models can. Its visual capability is code-based — SVGs, diagrams, interactive prototypes — plus much-improved image understanding (Opus 4.7 reads screenshots/mockups far more accurately than prior models). Any community workaround piping Claude into FLUX/Stable Diffusion via MCP is third-party, not an Anthropic feature.
Claude Design (launched April 17, 2026, Anthropic Labs) is the most directly relevant launch for this topic: a conversational visual-creation tool — separate from Artifacts and Claude Code — for prototypes, slide decks, one-pagers, and marketing visuals. Anthropic explicitly named product managers sketching feature flows as a target user, alongside designers and founders. It auto-detects brand/design-system context from an existing codebase, supports inline commenting, org-scoped sharing, and hands off directly to Claude Code for implementation. Anthropic reported 1 million+ users in its first week, and a Brilliant product designer reported pages that took "20+ prompts in other tools" took only 2 prompts in Claude Design.
Official Figma integration: a Figma MCP connector lets Claude read a Figma file's components/tokens/layout and turn them into production code, and a January 2026 Claude-FigJam integration lets users turn a Claude conversation into an editable FigJam diagram (user flows, architecture diagrams).
Official Adobe partnership ("Claude for Creative Work," April 28, 2026): connectors spanning 50+ tools across Photoshop, Illustrator, Premiere, Firefly, plus Affinity, Blender, SketchUp, and Ableton — Claude decides which tool to invoke and in what sequence for a stated goal. Anthropic's framing: "Claude can't replace taste or imagination, but it can open up new ways of working."
Real-world case study: a product designer at Jane Street published a detailed account of designing a feature by building a working Claude Code prototype directly instead of Figma mockups — with an honest caveat that fully-built prototypes can reduce collaborative iteration with engineers because the design already arrives "baked."

How AI is reshaping the design profession

Adoption has flipped fast: 91% of designers now use AI weekly (up from 54% a year earlier), and the average designer's AI toolstack grew from 3 to 7 tools. Notably, Claude (78%) now edges out ChatGPT (65%) among designers, and 65% use Claude Code specifically.
There's a real adoption gap vs. developers: only 31% of designers use AI for core design work vs. 59% of developers using it for core dev work — and designers report lower satisfaction (69% vs. 82%) and trust the output quality less, with inconsistent/unreliable AI output cited as the top barrier by 62% of designers.
The job market has genuinely polarized, mirroring the pattern seen in PM roles: UX job postings are down 71% from their 2022 peak, one in three organizations cut UX staff in 2024 (NN/g's worst year on record), and design's share of new hires at Big Tech dropped from 14% to 7%. At the same time, design job postings at AI-native/startup companies rose ~60% in 2025 — for fewer, more senior, more AI-literate hires, not junior production roles. The framing one industry piece used: "Product designers are not being replaced. They are being sorted."
What's emerging as irreplaceable: design leaders consistently point to taste, judgment, and the "fine-grain tuning" of the last 20% that AI can't close on its own. DoorDash's Head of Consumer Experience Design put it directly: AI gets you 80% there, but "fine grain control is the biggest gap." Nielsen Norman Group's 2026 outlook names "curated taste, research-informed contextual understanding, critical thinking, and careful judgment" as the durable core.
New roles and skills: titles like "design engineer" and "AI design expert" (Nike has created a dedicated Generative AI Design Expert role) are emerging, and 50% of design leaders now explicitly weigh AI fluency in hiring — though only 28% of companies have actually formalized evaluation or compensation changes to match, suggesting organizations are adopting the tools faster than they're updating how they manage the people using them.

What this means specifically for PMs ("PM-as-builder")

This is a real, named trend, not a hypothetical. PMs at Meta's Superintelligence Labs reportedly "vibe-code" prototypes and demo them directly to leadership instead of using slide decks, and Meta has added a live prototyping round to PM interviews.
LinkedIn replaced its Associate Product Manager program with an "Associate Product Builder" track in early 2026 — applicants must show a product they've already built and be comfortable with code/prototyping/AI-assisted workflows, with no formal product experience required. This is the clearest concrete signal that "can prototype with AI" is becoming an actual PM hiring bar, not just a nice-to-have.
Survey data backs up the trend: among PMs, the gap between current use of AI for prototyping (19.8%) and desired future use (44.4%) is the single largest swing measured across any PM AI use case — bigger than the PRD-writing gap.
This changes PM-designer collaboration, with real friction. The same survey shows designers have an even bigger prototyping-desire swing, and explicitly frames PMs as "increasingly encroaching" on design/engineering territory. One proposed fix from a Lenny's Newsletter contributor: designers should own and maintain the shared component library/design tokens that PMs draw from, so a PM's fast internal prototype doesn't get mistaken for a polished, designer-owned artifact.
The new core PM skill isn't coding — it's reviewing. Product strategist Paweł Huryn argues the skill that matters is reviewing an AI-built artifact against intent, risk, and quality — not reading the underlying code. PMs "don't have to learn to code," but they do need to know what "good enough to ship" looks like.
Real risks if PMs lean on this without guardrails: AI-generated prototypes create "the illusion of completeness" — visual output can look ~80% done while the underlying code is closer to ~20% production-ready, which sets unrealistic engineering expectations once a stakeholder has seen a working demo. Common, documented failure modes include accessibility gaps (AI defaulting to non-semantic markup), design-system inconsistency (ignoring existing tokens/components), and security gaps in fast-shipped AI-generated code.

Implications for PMs / Practitioners

You can now prototype your own ideas before involving design or engineering — but treat the output as a conversation-starter, not a spec. The LinkedIn APB program and Meta's vibe-coding norm show this is becoming an expected skill, not a shortcut to skip design entirely.
Loop design in before a prototype becomes "the plan." The most consistent failure mode in the research is a PM-built prototype arriving fully baked and engineering just running with it — bypassing accessibility, brand consistency, and design-system review. Decide up front whether a prototype is for internal exploration or stakeholder-facing, and treat those differently.
Learn to review, not just to prompt. The valuable new skill is judging whether an AI-built artifact matches intent and is safe to ship — the same shift happening in design leadership, where judgment is becoming more valuable as production gets easier.
If your product touches visual/creative work, know what Claude can and can't do. It doesn't generate images natively — don't promise a stakeholder image-generation capability Claude doesn't have. Its real strength for PMs is code-based prototyping (Artifacts, Claude Code, Claude Design) and reading/understanding visual material (screenshots, mockups) far better than before.
Watch the design org's polarization pattern as a preview of what's coming to PM roles. Senior, AI-fluent design hiring is up; junior/coordinator-style design hiring is down. The PM equivalent (per earlier research on AI agents and PM roles) is following the same shape — invest accordingly.

Sources

Note on sourcing: a handful of figures (MCP's "97M downloads/month," the "$80B combined cloud spend through 2029," and a few job-posting statistics) come from single secondary/aggregator sources without independent confirmation — treat these as directional rather than precise. The June 12, 2026 report of a possible export-control suspension on Fable 5/Mythos 5 access is very recent and may have changed by the time you read this — worth a fresh check before citing it publicly. The "Code to Canvas" Figma feature name was found only in a secondary blog, not confirmed directly on anthropic.com or figma.com.