Someone types a question into Perplexity, gets a clear, well-sourced answer back, and never learns which model produced it. Not because the information is hidden on purpose, but because the question of “which model” has quietly stopped being the point. A year or two ago, picking an AI tool meant picking a model: this one for coding, that one for writing, another for research. Increasingly, the tool itself makes that choice, invisibly, before the answer ever reaches the person who asked.

The second tier of challengers

Most of the attention in closed models goes to the four labs everyone can name, but a second tier is worth tracking too, because it shows how differently companies are choosing to compete. Amazon has its Nova line, built less to top leaderboards than to be the default model wired into AWS and Alexa, wherever Amazon already owns the customer relationship. Microsoft has been developing its own MAI models, a hedge against relying solely on its partnership with OpenAI, and a way to control more of the stack that Copilot and Azure sit on. Mistral, a company often filed under “open models” because of its open releases, also sells closed, higher-end models to enterprise customers who want stronger performance than the open versions offer and are willing to pay for it. None of these three is trying to be the single best model in the world. Each is trying to be good enough, embedded in a business relationship, and available on terms that make sense for whoever already depends on that company for something else.

How routing actually works

The more structurally important development, though, isn’t a new entrant building its own frontier model. It’s platforms that build no frontier model at all and instead sit on top of many. Perplexity is the clearest example: it takes an incoming question, evaluates what kind of task it actually is (a quick factual lookup, a multi-step research question, a piece of code), and sends it to whichever underlying model is best suited for that specific job, sometimes a model from one lab, sometimes another, depending on the request. For a simple question this happens once. For a complex, multi-step task, some of these routing systems now coordinate dozens of different underlying models within a single session, each handling a different piece of the work, none of it visible to the person typing. The protocol work described in MCP: The Standard Becoming the USB of AI is part of what makes this possible: a shared way for a system to call out to different tools and models without custom wiring for each pairing is what lets a router treat “which model handles this” as a decision it can make on the fly rather than a choice baked in at build time.

Once a router is choosing on someone’s behalf, the brand name on the underlying model stops mattering to the person using it. The user experience is the router’s experience: its interface, its speed, its judgment about which model to call. Power shifts away from whoever builds the best individual model and toward whoever controls the routing decision itself, because that decision point is where the user’s attention actually lives. This is close to what search engines once did to individual websites. A website could be the best possible source for a given question, meticulously written and technically excellent, and it still only reached someone if a search engine decided to surface it; the search engine, not the website, controlled whether that quality ever got seen. Increasingly, a router plays the same role toward models: it decides which model you’ll even talk to, and the model, however good, only matters if the router picks it.

That is close to the opposite bet from the one described in OpenAI / GPT: The Distribution Advantage, where a lab wins by putting its own named model directly in front of an enormous audience and making sure people know whose product they’re using. A router does the reverse: it deliberately obscures whose model answered, because the router’s value proposition is that the person asking shouldn’t have to know or care. Both are viable strategies. But only one of them keeps the model’s name on the label.