Frontier APIs are the quickest way to find out what AI can do. Once a workflow runs at scale, the value moves to the learning loop around the model, and open-weight models let you own that loop.
What changes at production scale.
Frontier APIs drove the first wave of adoption because teams could test a capability and ship it without building infrastructure. They are still the fastest route to discovery. Building is a different job, though, and sharing the same frontier intelligence as every competitor wears down long-term advantage.
Open-weight models have also caught up. GLM-5.2, which Z.ai released in June 2026 with open weights under an MIT license, scores within a few points of leading closed models on the coding benchmarks Z.ai published. With models like it available, teams ask practical questions about moving a workflow over, confirming quality, controlling cost and avoiding lock-in to one vendor.
At scale, the test is no longer whether a model can do the task. It is whether the model is fast, affordable, dependable and tailored enough to get better with each interaction.
What matters under real volume.
Take an AI support product. Early on, the open question is whether a model can resolve tickets at all, and a frontier API answers it well. Under real volume, other things matter: recurring failure modes, the human corrections that improve outcomes, and the customer-specific policies that apply. The value now sits in the system built around repeated answers.
Volume changes the math too. A frontier model may be right for occasional heavy reasoning. A narrow prediction that runs millions of times a day puts latency, consistency and unit cost first, and the best fit is often a model adapted to that one task.
Internal copilots work the same way. Hand evaluation and improvement to an outside model and you build a dependency, not an asset, because the traces and corrections that should improve your product are never kept.
Paying for intelligence twice.
Open-weight models let a company keep its traces, adaptations, evaluations and deployment choices. That is how model usage becomes proprietary advantage.
Microsoft CEO Satya Nadella named the underlying problem the "Reverse Information Paradox" in an essay he posted on X in July 2026: "You essentially pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it!"
The more a system depends on private context and company-specific standards, the more of its value lives in the learning loop. Lose control of that loop and you pay in both cash and knowledge.
Where Orchestra fits.
AI's first phase showed that models can do useful work. The next one settles who keeps what that work teaches. Leading teams will write their own evaluations, record failures, route requests well and adapt from real usage while holding costs down.
Orchestra is built for that shift. It turns production traces into evals, shows which tasks need a frontier model and which suit a specialist, and trains specialist models your team owns, so the learning stays with your team.