For a few years the argument seemed settled. The most capable artificial-intelligence models would be built behind closed doors, by a handful of well-funded labs, and rented out through an interface that revealed nothing of the machinery inside. Open alternatives — models whose weights anyone could download, inspect and run — were treated as a worthy hobby, perpetually a generation behind, useful for tinkering but unserious for real work. That consensus has quietly collapsed. By mid-2026, the gap between the best open-weight models and the proprietary frontier has narrowed from a chasm to a margin, and the entire industry is rearranging itself around the new geometry.
The closed giants still hold the very top. For the hardest reasoning, the longest context and the most reliable performance, the proprietary frontier labs remain ahead, and they intend to stay there. But the distance has shrunk to the point where, for the overwhelming majority of practical tasks, an open model that costs a fraction as much to run is simply good enough — and "good enough, far cheaper, fully under your control" is one of the most persuasive sentences in enterprise software.
The case on each side
The argument for the closed approach is coherent and not merely self-serving. Frontier capability is genuinely expensive to produce, and someone has to pay for the compute and the talent; a subscription model funds the next leap. Concentration also makes safety tractable, the labs argue: a model you control through an interface can be monitored, rate-limited and switched off if it misbehaves, whereas a model whose weights are loose in the world can never be recalled. Keep the most powerful systems behind a controllable boundary, the reasoning goes, and you keep a hand on the brake.
The open camp answers with a different set of values. A model you can download and run is a model no vendor can deprecate, price-gouge or quietly alter beneath you. You can audit it, fine-tune it on your own data, run it inside your own walls and know that nothing leaves the building. And the safety argument cuts both ways: open weights mean thousands of independent researchers can probe a model for flaws rather than trusting a single company's private assurances. Control, to this camp, is not something you cede to a vendor; it is something you keep.
"The question stopped being 'which model is best.' It became 'which model do I actually own' — and for a lot of companies that changes the answer."
A machine-learning lead at a European bank — interviewed for this articleEnterprises hedge by running both
What is striking, talking to the people who actually deploy these systems, is how few of them treat it as a binary. The sophisticated answer in 2026 is not open or closed but open and closed, routed by task. The pattern recurs across industry after industry: a closed frontier model for the genuinely hard, low-volume problems where capability justifies the price and the dependency, and an open model running in-house for the high-volume, privacy-sensitive, cost-conscious work that makes up the bulk of the load.
This hedging is partly a negotiating tactic. An enterprise that can credibly run open models in production has enormous leverage over its closed-model vendor; the threat of migration is no longer theoretical. It is partly resilience — no single supplier can hold a critical workflow hostage. And it is partly simple economics: for a task you run a million times a day, the difference between a frontier model's per-token price and an open model's near-zero marginal cost is the difference between a viable product and a money pit. The result is an architecture built deliberately to avoid lock-in, a posture this same impulse toward independence echoes far beyond software, as our investigation into Europe's decade-long sovereignty bet describes.
The European angle
Nowhere is the open-model case more politically charged than in Europe, where it has fused with the continent's broader anxiety about technological dependence. For policymakers in Brussels, a future in which every important AI system is rented from a handful of non-European companies looks uncomfortably like the chip dependency they are spending hundreds of billions to escape. Open-weight models offer an appealing alternative: a capability that can be hosted on European soil, audited by European institutions and adapted to European languages and rules, without a permanent tether to a foreign vendor's pricing and policies.
That has made open models something close to an instrument of industrial strategy. European research groups and startups have leaned into open releases not only on principle but as a competitive wedge — a way to matter in a field whose frontier they cannot yet out-spend. It is the same logic running through the continent's semiconductor push: you may not own the absolute leading edge, but owning a credible, controllable, home-grown capability is its own kind of power.
None of this means the closed labs are losing. They are still defining what is possible, still capturing the most lucrative high-end work, still a generation ahead at the very top. But the war the headline promises is real, and it is being fought less over who builds the single best model than over the shape of the whole market — whether intelligence becomes a utility rented from a few towers, or a commodity that anyone can run. In mid-2026, both futures are still live, and the smartest players are quietly betting on both at once.