Anthropic has crossed a line this week that signals a fundamental shift in how AI labs operate: it launched Claude Design, a product that directly competes with its own paying customers. On the same day Anthropic announced Claude Design, Figma's stock dropped 7.28 percent, and board member Mike Krieger resigned. This isn't coincidence. It's a preview of a much larger problem facing companies that have bet their roadmaps on foundation models .

Why Are AI Labs Suddenly Building Products?

For the past two years, the story was simple: foundation model companies like Anthropic and OpenAI made models, and everyone else built products on top of them. That division of labor is breaking down. Anthropic shipped Opus 4.7 this week and simultaneously announced that an even better model, called Mythos, remains locked inside Project Glasswing and isn't available for purchase. OpenAI gated two additional models, GPT-5.4-Cyber and GPT-Rosalind, behind restricted access programs. Meanwhile, Canva announced a complete rebuild of its platform around an in-house AI model, powered by a conversational interface .

The pattern is clear: the best AI models aren't for sale anymore. Instead, they're reserved for the labs' own products. A year ago, the most capable model any lab had was the one you could sign up for. That stopped being true this week .

What Does This Mean for Companies Built on Claude?

If you've built a product or service on top of Claude, Anthropic's new strategy creates a strategic vulnerability. The company that supplies your core technology is now also building competing products. Figma learned this the hard way. The design platform had integrated Claude's capabilities into its workflow, but Claude Design now offers similar functionality directly to end users, bypassing Figma entirely .

This creates a cascading problem. Companies that depend on frontier AI models now face a choice: diversify their model dependencies, invest in proprietary data and workflows that create defensibility beyond the model itself, or accept that their supplier may eventually compete with them. Anthropic's expansion this week included not just Claude Design, but also Claude for Word, a 4x expansion of its London office, and Claude Code Routines. It's no longer accurate to call Anthropic a foundation-model provider. It's a product company that also makes models .

How to Protect Your Business From AI Supplier Risk

• Build a Proprietary Moat: Distribution, workflow depth, and proprietary data are now more valuable than the underlying model. If your only advantage is that you wrap Claude in a user interface, you're vulnerable. Companies like Canva are responding by building their own models, while others are doubling down on domain-specific data and workflows that competitors can't easily replicate.
• Diversify Your Model Dependencies: Relying on a single frontier lab is increasingly risky. Companies should test their hardest prompts and workflows against multiple models, including Anthropic's Claude, OpenAI's GPT models, and open-weight alternatives. This isn't just about redundancy; it's about negotiating leverage.
• Plan for Competitive Scenarios: Spend time this week working out what you'd do if Anthropic or OpenAI shipped a product that directly competes with yours. What's your fallback? What's your differentiation? This exercise should inform your product roadmap and investment priorities.

The broader implication is sobering: every time a frontier lab ships a full-stack product, it starts a "what's our Gemini plan?" conversation somewhere in the industry. Companies are quietly diversifying away from single-supplier dependencies .

Are the Economics of AI Production Sustainable?

There's another pressure driving this shift: cost. Uber's Chief Technology Officer told staff this week that the company had already burned through its entire 2026 AI budget by April, mostly on Claude Code. Anthropic responded by quietly reducing Claude's default effort level to "medium," a change caught by an analysis of 6,852 sessions. OpenAI killed Sora, its video generation model, which was reportedly costing $1 million per day in compute. Perplexity put its Personal Computer agent behind a $200-per-month paywall .

Production AI now behaves like cloud compute, not Software-as-a-Service (SaaS). The economics are brutal: running sophisticated AI agents at scale is expensive, and the cost structure doesn't match traditional software pricing. This is forcing labs to make hard choices about which products to keep, which to gate behind premium access, and which to build in-house where they can control costs .

What About AI Memory and Long-Running Agents?

One of the infrastructure problems driving costs is context management. AI models like Claude Opus 4.7 have a 1 million token context window, which can accommodate roughly 555,000 words or 2.5 million Unicode characters. Claude Sonnet 4.6 also has a 1 million token context window but holds roughly 750,000 words or 3.4 million Unicode characters because it uses a different tokenizer. However, the actual usable context space is often 10 to 20 percent less when you account for system prompts, tools, custom agents, and other overhead .

Cloudflare addressed this problem this week by launching Agent Memory, a managed service that stores AI conversation data off to the side and recalls it when needed. The service allows AI agents running for weeks or months against real codebases to maintain useful memory without bloating the context window. Tyson Trautmann, senior director of engineering at Cloudflare, explained the value: "Agents running for weeks or months against real codebases and production systems need memory that stays useful as it grows, not just memory that performs well on a clean benchmark dataset that may fit entirely into a newer model's context window" .

Cloudflare

"Agent Memory is a managed service, but your data is yours. Every memory is exportable, and we're committed to making sure the knowledge your agents accumulate on Cloudflare can leave with you if your needs change," stated Tyson Trautmann, senior director of engineering at Cloudflare.

Tyson Trautmann, Senior Director of Engineering, Cloudflare

The service is currently in private beta and accessible via Cloudflare Workers or REST API. This kind of infrastructure maturation is necessary if AI agents are going to be viable in production, but it also adds another layer of cost and complexity to the stack .

What Should You Do Right Now?

The immediate takeaway is this: if you've built on top of a frontier lab's model, your competitive advantage is no longer the model itself. It's what you do with it. That means investing in proprietary workflows, domain-specific data, and defensible distribution. It also means testing your critical workloads against multiple models and having a plan for what happens if your primary supplier launches a competing product .

For companies considering AI infrastructure investments, the lesson is equally clear: the era of simple model-as-a-service is ending. The future is more complex, more expensive, and more competitive. But it's also more defensible if you build the right moat.