A thousand models, one question

Choice was supposed to be the prize of an open AI ecosystem. Without the right infrastructure, it is a burden.

In 2004, the psychologist Barry Schwartz described the paradox of choice: past a certain point, more options make decisions harder and outcomes worse. Two decades later, AI has built a near-perfect demonstration. There are now more than a thousand capable models across text, image, audio, and video, with new releases weekly. Ask which is best and the honest answer is a question back: best at what?

The old shortcut, just use the frontier model, is losing its logic. The frontier is converging. In a single year, the performance gap between the best model and the tenth-best roughly halved, and the leading two finished almost level. The strongest model overall is rarely strongest at everything. One leads on code, another on long documents, another on image generation, and the rankings change monthly.

No person, and no business, can track this by hand. Which is why the answer to abundance is not a better leaderboard. It is routing: infrastructure that decides, task by task, which model should run.

Routing is a complex problem, and the field is still experimenting. But some fundamentals are starting to be established. A mature system weighs the nature of the task, the price of each model, and whether part of the answer already exists and can be reused, then matches the work accordingly. Simple work goes to the cheapest capable model, and only the hard work escalates. A frontier model might plan a task while a lighter one executes it. And the emerging consensus among the most advanced adopters is that people should not be picking models at all. That judgement belongs to the system.

The results are measured, not asserted. In peer-reviewed work from Berkeley, a trained router cut costs by more than 85% on a common benchmark while holding around 95% of the strongest model's quality. In industry, companies such as Coinbase have built internal gateways that route work across models automatically, reporting spend cut nearly in half while usage grew. Every business that builds its own version proves the same point: routing is no longer optional. It is becoming the architecture of how AI is used.

Doing it well at scale, across every model and modality, for millions of users rather than one engineering team, is a far harder problem. It requires enormous amounts of data and continuous innovation, because the models, their prices, and their strengths change constantly. That is the problem Smart Routing exists to solve.

Choice stops being a burden when something chooses well on your behalf. The durable advantage is not owning the best model. It is combining many models well. That is how a thousand options become one good answer.

SOURCES

Barry Schwartz, The Paradox of Choice (2004). Stanford HAI, 2025 AI Index Report. Ong et al., RouteLLM: Learning to Route LLMs with Preference Data, ICLR 2025. Brian Armstrong (Coinbase), X, June 2026.