Most conversations about enterprise AI still revolve around the question of ‘which model wins?’ People compare benchmarks, refresh leaderboards, and judge every new release on whether it edges past the last. That focus is understandable, but it is increasingly the wrong one.
The decision that matters more is far less visible: who, or what, is routing work between models. As the market moves from ever-larger models towards cheaper, smarter systems, value is migrating to the layer that decides which model handles which task, and to the commercial and governance choices that keep that layer flexible.
Routing, not model size, is the real product
The hardest part of deploying smaller, task-specific models into sensitive, live enterprise environments is not the small model itself. It is the routing layer that sits on top of it. As Perplexity CEO Aravind Srinivas has pointed out, the real product is the system that decides which model handles which task.
That decision cannot be pushed onto people. Give a user the choice between a small, medium, and large model, and the behaviour is entirely predictable: they pick the large one every time, drop to the medium only when their token budget gets tight, and never touch the small one. Making that judgement call dozens of times a day is simply too daunting.
The routing therefore has to be automatic, and it has to be right. In a live, sensitive enterprise environment, the bar is something like 99.9% accuracy. Fall below that, and the first time a critical task is routed to the wrong model, trust disappears, and frustrated users switch the whole thing off. In this context, reliability is not just a feature, it is the entire value proposition.
The open-weight switch is already underway
Alongside the routing question, the economics of the model market are changing. Open-weight models, whose trained parameters are publicly released, are improving quickly, and enterprises are already starting to reduce their reliance on premium closed-source models.
By this year, open-weight models had reached close to parity with the frontier. Once outputs are genuinely hard to tell apart, economics take over. If an alternative is up to three times cheaper, switching stops being a risk and becomes the default.
The leading indicator to watch is the tooling. The first real sign of this switching wave is the adoption of alternative agent harnesses, such as opencode, which make swapping models trivial. Combine easy model-switching in the harness with wide availability of open-weight models hosted in the wild at attractive prices, and the switches will happen quietly, wherever output quality is indistinguishable.
That quietness is worth noting. Rather than arriving as large, announced migrations, this shift is more likely to show up as a series of incremental decisions, taken task by task, wherever the quality gap has closed.
Flexibility is now a governance issue
So how should organisations balance cost, performance and governance when selecting AI models? The single most important principle today is staying flexible. The market is moving too fast for any organisation to lock itself in.
That flexibility has a human dimension that is often underestimated. End users get attached to their desktop app; developers get attached to their harness. These are the surfaces where switching either stays easy or quietly becomes impossible, so it pays to offer flexible desktop apps and agent harnesses early. Lock-in, in other words, can set in at the application and harness layer, not just at the model.
Commercial terms deserve the same discipline, and the rule is to keep commitments short. For most organisations, monthly is already the longest commitment they should be comfortable with: long enough to be practical, short enough to change course when a cheaper or better option appears.
Governance is where data residency comes in. Today, residency can genuinely restrict options: a given model may be off the table simply because it is not hosted in the required region. But that constraint is likely to ease quickly. Price pressure and demand are already driving open-weight hosting into all the major regions, across the US, the EU and Asia, so residency becomes less of a hard limit and more of a short-term consideration.
Asking the right question
Taken together, these shifts point to a different set of priorities for enterprise technology leaders. The question is no longer which model to bet on, but whether the organisation has built the conditions to move between models as the market changes: routing that is automatic and reliable enough to earn trust, tooling that keeps switching easy, and commercial and governance arrangements that leave room to change course.
The model race will keep generating headlines, and today’s leader may not be tomorrow’s. With the market moving this fast, the advantage will go to the organisations that stay flexible enough to switch whenever a cheaper or better option appears.