The core difference: do you get the model, or just access?
Proprietary models like GPT, Claude, or Gemini you use exclusively via the provider's interface – you never get the model itself, you send your request and get an answer back. Open models like Llama or Mistral publish their "weights" (the trained model): you can download them and run them yourself on your own infrastructure. That's the real difference – not the quality of the answers, but who holds the model.
"Open" is a spectrum
Important for context: "open" isn't a clean yes/no line. Many models publish the weights but not the training data – and the license can restrict commercial use. So "open weights" doesn't automatically mean "free for anything." With the term "open source," it's always worth checking the specific license.
What that means in practice
Data sovereignty & control
An open model you can run in-house – the data then never leaves your infrastructure (see modules 7 and 20). With proprietary models, the inputs go to the provider, with the familiar conditions (DPA, training opt-out – see module 7).
Cost
Proprietary means: pay per use, no infrastructure of your own required. Open means: no per-request license fee, but costs for hosting and the know-how to run the model at all.
Performance & effort
The absolute top tier is often proprietary, and they're turnkey. Open models have caught up strongly, but running them demands real technical capacity.
Flexibility & lock-in
Open models avoid dependence on a single vendor and can be customized more deeply (including fine-tuning, see module 16). Proprietary is more convenient but ties you to one vendor.
Why this matters for you as a decision-maker
For most companies, a proprietary model via an interface is the pragmatic default: turnkey, top of the performance range, low effort. Open source gets interesting when data sovereignty is the top priority, when the technical capacity is in-house, or when very high request volume tips the economics. It's a question of fit, not ideology – and often the answer is a hybrid: proprietary for most of it, a self-hosted open model for the especially sensitive part.