What ‘open’ and ‘closed’ actually mean
A closed (proprietary) model — GPT-4o, Claude, Gemini — is accessed only through the provider’s API or app. You never see the weights; you send tokens and get tokens back. An open model — Llama, Mistral, Qwen, Phi — publishes its weights so anyone can download, run, and fine-tune it on their own hardware. Note the nuance: most “open” models are really open-weight, since the training data and full recipe are rarely released, and licences vary from fully permissive (Apache 2.0, MIT) to restricted. The choice between the two camps is one of the most consequential decisions when building on AI.
Performance and capability
The capability gap has shrunk fast. Top open-weight models now rival older frontier closed models on standard benchmarks and handle the large majority of everyday tasks — summarisation, drafting, classification, code assistance — extremely well. The leading closed models still hold an edge on the hardest reasoning, the longest context windows, and the most polished multimodal (image, audio, video) capabilities, and they ship improvements continuously without any work from you. For frontier-grade results with zero infrastructure, closed still leads; for most production workloads, open is genuinely good enough.
Privacy, control, and cost
Privacy and control favour open models: you can run them entirely inside your own environment, so sensitive data never leaves, you control exactly which version runs, and no provider can deprecate the model out from under you. Cost and convenience usually favour closed models: you pay per token with no GPUs to manage, which is cheaper at low or bursty volume and trivial to start. Self-hosting an open model only wins on cost at high, steady throughput where you can keep expensive hardware busy — and it adds real engineering overhead. Closed enterprise tiers also offer no-training and data-residency contracts that satisfy many privacy needs without self-hosting.
A decision guide
Choose a closed model if you want the best possible quality with no infrastructure, your volume is low or unpredictable, and an enterprise no-training contract meets your privacy bar. Choose an open model if data must stay in your environment for regulatory reasons, you need full control over the model and its lifecycle, you have high steady volume that justifies hardware, or you want to fine-tune deeply on your own data. Many mature teams do both — a closed API for the hardest tasks, a self-hosted open model for high-volume or sensitive ones — and always check the specific licence before shipping a product.