Open models publish their weights. You can run them yourself, fine-tune on proprietary data, apply quantization to cut infrastructure costs, and avoid vendor lock-in entirely. Closed models are API-only: you get frontier capabilities with no infrastructure burden, but you're dependent on the provider's roadmap, pricing, and continued existence.

Open-weight models hand you control and cost leverage at the price of operational load. Your team manages deployment, scaling, and updates. Closed models eliminate that burden and tend to move faster on capability and safety improvements — but the dependency is real and compounds over time as you build around their APIs and data formats.

The pragmatic approach is almost always mixed. Use closed models for prototyping, capability-heavy tasks, and anything where time-to-value outweighs cost. Use open models for high-volume or regulated workloads where inference cost and data residency dominate. This is a direct application of the build-vs-buy framework — no universal right answer, but a clear set of variables to evaluate per use case.

One thing worth watching: the capability gap between open and closed models has been narrowing faster than most expected. The assumption that frontier intelligence requires a closed API is less true every six months.