Open models are catching the frontier. Here is what that means for your business.

Two years ago the assumption was simple. Intelligence was something you rented from a handful of American labs, paid for by the token, and accessed by sending your data off to their servers. That assumption is quietly breaking. Open-weight models, the kind you can download, run on your own infrastructure, and own outright, have closed most of the gap to the frontier over the course of 2026. For a business, that is not a technical curiosity. It changes what is possible and what it costs.

The gap is now months, not years

The open models coming out of labs like DeepSeek, Alibaba's Qwen family, and Meta's Llama line have caught up faster than almost anyone predicted. On coding, which is the most common serious business use, open models now match the best closed ones on the standard industry benchmarks. On reasoning and long documents they are close behind. The frontier still leads on the very hardest problems, but the distance is now measured in a few months of release cadence rather than a generation. By some estimates the best open weights trail the closed frontier by only a matter of months.

For the vast majority of business problems, the question is no longer whether an open model is good enough. It is whether you need the frontier at all.

Why this matters for cost and control

There are two practical consequences, and both favour the business. The first is cost. Open models under permissive licences run far cheaper per token than frontier APIs, often by a factor of ten or more, and the price keeps falling as hosting improves. The second is control. A model you run yourself keeps your data on infrastructure you choose, cannot be deprecated out from under you, and cannot have its terms or its price changed by a vendor you do not control.

• Cost per token on open models has fallen by an order of magnitude, and often more, against frontier APIs.
• Your data stays on your own infrastructure instead of being sent to a third party.
• No single vendor can raise prices, tighten limits, or retire the model your systems were built on.
• You can fine-tune an open model on your own operation, something most frontier APIs do not allow.

When the frontier is still worth it

This is not an argument that open always wins. The frontier models still lead on the hardest reasoning, the most complex multi-step agent work, and the newest capabilities in their first months of life. The right answer for a business is almost always a mix: the frontier where the problem genuinely demands it, and open weights for the large volume of ordinary work that does not. Knowing which model to point at which problem, and being willing to change your mind as the models change, is the actual skill.

The supplier who only sells you frontier tokens has a reason to tell you the frontier is the only serious option. The honest answer is that the map has changed. The cheapest and most private option is now good enough far more often than it was a year ago, and a business that understands that pays less and owns more.