Open-weight vs closed AI models: what the difference means for you

A closed model (GPT-6 Astra, Claude Fable, Gemini) is reachable only through its vendor's API. An open-weight model (Meta's Llama family, Mistral's open releases, DeepSeek, Google's Gemma, OpenAI's gpt-oss) publishes the trained parameters so anyone can download, run, modify and redistribute it under a licence.

"Open-weight" is not the same as "open source": the weights are public but the training data and full recipe usually are not.

Why companies choose open weights

• Control. Run it in your own data center or cloud account. No data leaves. No vendor can change or retire the model under you.
• Cost at scale. Above a certain volume, renting GPUs and running an open model is cheaper than paying per token.
• Customisation. Fine-tune on your own data without restrictions.
• No rate limits or outages you do not control.

Why companies choose closed

• Capability. The top closed models still lead on the hardest reasoning and agentic tasks, though the gap narrows every few months.
• No operations. No GPUs to provision, no serving stack to maintain, no security patches for inference servers.
• Features. Tool use, caching, batch, long context and safety tooling arrive first on the vendor API.

The security angle

Open weights can be audited and run air-gapped, which regulated industries like. They can also be fine-tuned by anyone to remove safety behaviour, which is why dual-use capability such as advanced cyber-offence tends to ship closed and gated (OpenAI's GPT-6 Cyber is available only to vetted customers).

A practical rule

Prototype on a closed API for speed. Once a workload is stable and large, price out an open model on your own infrastructure for that specific task. Many companies end up with both: closed for the hard, low-volume tasks and open for the high-volume, well-defined ones.