The case against one big model

Simplicity is appealing in tooling. One provider, one interface, one model to understand and work with. When that model is good, the appeal strengthens — why reach for variety when the thing in front of you is already capable?

This is a reasonable heuristic for individual sessions. It is a fragile strategy for anyone building on AI capabilities or relying on them for serious work.

The monoculture problem

The analogy that keeps coming back to us is agriculture. Industrial monocultures are extraordinarily productive — until a blight arrives. The efficiency of the monoculture and the uniformity of the monoculture are the same property. When conditions change, diversified systems degrade gracefully. Uniform systems fail fast.

AI model capability follows a similar pattern. Every model has a characteristic profile: tasks it handles with confidence, tasks where it hedges, tasks where it silently underperforms in ways that are hard to detect. When you route all your work through a single model, you inherit that profile completely — the strengths and the blindspots, without the ability to distinguish between them.

A platform that only knows how to ask one model is not just dependent on that model performing well. It is dependent on that model's weaknesses never being the weaknesses that matter on the day you need them not to matter.

Capability is not uniform

There is a tempting assumption built into the phrase "best model": that capability is a single dimension, and the best model is best along all of it. That is not how this technology works.

Different models carry different training emphases, different reasoning approaches, different performance profiles across domains. The model that is most capable on complex legal reasoning may be different from the model that is best at structured data analysis, which may differ again from the model best suited to generating drafts with a specific register. "Best" is always best-for.

This is not a temporary condition while the field matures. It is structural. The diversity of tasks people bring to AI platforms exceeds what any single model's training can optimise for uniformly.

What variety actually buys

Routing work to the right capability — rather than the available capability — is what separates a capable tool from a reliable one. It also changes the relationship between the platform and the model landscape. When your workflow depends on one provider, model deprecations, pricing changes, and service disruptions become your disruptions. When the platform can route across options, those events are absorbed before they reach you.

There's a second benefit that's harder to articulate but equally real: exposure to a range of model outputs builds calibration. If you've only ever seen how one model approaches a question, you don't know how much of what you're reading is the answer versus the model's characteristic voice and blind spots. Variety makes the underlying uncertainty more visible, which makes the work more honest.

See /models for the current range of capabilities available across the platform.

The practical argument

We are not making a claim that more is always better, or that every task needs multiple models in competition. Most of the time, one well-chosen approach is exactly right. The argument is narrower: that choosing to be unable to use anything except one model is a structural liability, not a simplification.

The platforms that serve professionals well over time are the ones that kept options open. The ones that bet everything on a single capability source have, historically, either renegotiated or rebuilt — both of which are expensive, and both of which happen at the worst possible moment.

Simplicity in an interface is a virtue. Simplicity in your capability foundation is a risk. They are not the same thing, and it's worth holding them apart.