For years the pattern held: closed frontier labs set the state of the art, and open models followed a year or two behind. In 2026 that gap narrowed sharply — and on the metric that matters most for deployment, cost per useful token, open models often win outright.
What changed
Three forces converged. Architectures matured — nearly every serious open release now uses Mixture-of-Experts, buying frontier-class capacity at a fraction of the active compute. Training recipes spread quickly, so hard-won techniques no longer stay proprietary for long. And a competitive field of labs kept shipping strong open-weight models with permissive terms.
The result: models you can download and run now handle the majority of real workloads well enough that the marginal quality of a closed model doesn't justify its price for most tasks.
The question flipped. It's no longer "is the open model good enough?" — it's "does this task actually need the closed one?"
Why cost decides
For high-volume production — support, extraction, classification, routine generation — the economics are decisive. An open model you host is cheaper per call, keeps data on your own hardware, and can be tuned to your task. Closed frontier models remain the choice for the hardest reasoning and longest-horizon agentic work, but that's a shrinking slice of total volume.
The emerging default
The winning pattern is a cascade: a capable open model handles the common cases, escalating only genuinely hard requests to a frontier model. You get frontier quality where it matters and open-model economics everywhere else. Teams that architect for this — rather than defaulting every call to the most expensive model — are quietly running circles on cost.
