Thursday, 2 July 2026

What strikes me about this week's model announcements isn't that there are more of them — there's always another LLM launch — but that they're revealing something fundamental about where the real value lies now. It's no longer in raw capability. It's in specificity.

Anthropic's Claude Sonnet 5 is competent, sure, with improvements in coding and reasoning. But the quieter story is Travelers building TravelersLLM, a model designed from the ground up for property and casualty insurance workflows. That's not flashy. It won't trend on social media. But it's the kind of thing that actually changes how insurance underwriters work — faster decisions, fewer errors, less time hunting through documentation. When you train a model on domain-specific data and patterns, you're not just getting better performance on a benchmark; you're building something that understands the peculiar logic of your business in a way a general-purpose model never will.

The European Commission backing the EUROPA consortium to build a 24-language LLM across EU languages points in a different direction, though it arrives at a similar conclusion. This is about sovereignty and accessibility. Open-source, European-built, multilingual — it's a deliberate counter to the concentration of AI capability in a few American labs. Whether it will genuinely compete on quality remains to be seen, but the bet here isn't purely technical. It's political and practical: can we build infrastructure that works for our economies and doesn't depend on American goodwill?

By the way, Base44's move to build its own model for web design rather than licensing someone else's output feels almost quaint in comparison. They're solving a real problem — the homogenized, template-y look of AI-generated websites — but it also hints at something worth watching. As foundation models mature and plateau in raw capability, the differentiation increasingly comes from either specialized training on your particular problem domain, or clever system design that combines models in novel ways. Base44 is doing the former. Others are experimenting with hybrid workflows that blend local models for privacy with cloud-based reasoning for scale.

This isn't to say general-purpose models are becoming irrelevant. Sonnet 5 and whatever Anthropic releases next will remain the backbone for many applications. But I think we're entering a phase where the smartest companies are asking a different question: not "which foundation model should we use?" but "what's the minimal capability we need from a general model, and how do we build the rest?" The answer increasingly involves domain-specific fine-tuning, specialized architectures, or carefully designed hybrid systems. The winners won't be the ones with the biggest model. They'll be the ones who understand their problem deeply enough to build something purpose-built.