Claude Cowork Bet on One Model. We Didn't.

Anthropic shipped Claude Cowork in January and spent the first half of the year turning it into a real product — enterprise connectors in February, and mobile and web access this month. It's a genuinely good agent. It reads your files, moves across your apps, and comes back with a finished deliverable instead of a paragraph of advice.

It also only runs on Claude. That's not a criticism — it's the point of the product. Anthropic builds the model and the wrapper, so of course the wrapper is a showcase for the model. But it means every task Cowork completes is capped by whatever Claude can do today. If a better model ships next quarter from anyone else, Cowork doesn't get better. It waits for Anthropic's next release.

The wrapper that doesn't care who wins

Compare that to the products built the other way around. Viktor — the Slack- and Teams-native AI employee that's grown to a $450M valuation off a $75M Series A in a matter of months — says it plainly on its own pricing page: model-agnostic, the right frontier model for each job. Perplexity Computer is built the same way: it routes a task to whichever model is best suited for it, and swaps that choice out the moment a better one exists.

Neither company is betting the product on one lab staying ahead forever. They're betting that no lab stays ahead forever — which, if you've watched this market for more than six months, is the safer bet by a wide margin.

Models leapfrog. Habits don't.

We wrote about this exact dynamic a couple months ago, and Cowork's rollout is making the argument for us in real time. The frontier changes hands every few months — a new release from one lab is the best model on the market for a quarter, then it isn't. That's fine if you're a wrapper. It's a structural risk if the wrapper and the model are the same company.

• Single-model products inherit the model's ceiling. Cowork is exactly as good as Claude is this week. No better.
• Model-agnostic products inherit the market's ceiling. Viktor and Perplexity Computer get better every time any lab ships something stronger — they just route to it.
• Switching cost compounds in the wrapper's favor. The team that's already living in your Slack, your files, your workflow doesn't lose you when a model regime changes. They just quietly upgrade the engine underneath you.

Anthropic being great at building models is not the same claim as Anthropic being great at building the product you open every morning. Those are two different businesses, and increasingly, two different winners.

Why we build this way too

We don't ship a single client on a model we can't swap out from under them. Every Appsurd build — including the agents we stand up for clients — is architected so the model is a config value, not a foundation. When a stronger or cheaper model ships, we route to it the same week, and the client never has to know or care that anything changed underneath.

That's the actual product decision hiding inside "model-agnostic." It's not a marketing line. It's an admission that nobody, including the labs themselves, knows which model will be best in six months — so the only durable move is to build the thing that doesn't have to guess.

Cowork will keep getting better. So will Claude. That's genuinely good for everyone building on top of it, us included. But the coworker that outlasts this cycle won't be the one wired to a single model — it'll be the one that was never betting on any one model to begin with.