What is Closed Weights? — AI Glossary

What it is: Closed-weight AI models are ones whose trained model files are kept private by the company that built them. You access them only through paid APIs — you can’t download or run them yourself. GPT-5, Claude, and Gemini are all closed-weight.
Who it is for: Anyone wanting to understand the business model behind major AI products, especially the difference from open-weight alternatives.
Best if: You’re evaluating AI vendors, comparing total cost of ownership, or just trying to understand why some models are downloadable and others aren’t.
Skip if: You’re a casual ChatGPT user who doesn’t care about the underlying business model. Want one practical AI workflow every morning? Subscribe to our free daily newsletter.

What are closed weights?

Closed-weight AI models are ones where the trained weights — the actual files that make the model work — are kept private by the company that built them. To use a closed-weight model, you call its API and pay per request. You never see or download the model files; you can’t run it on your own hardware; you can’t modify it directly (only through fine-tuning APIs the company provides).

All the frontier consumer AI assistants are closed-weight: ChatGPT (GPT-5 and below), Claude, Gemini (the paid versions of Pro and Ultra), Grok (mostly), Perplexity’s own Sonar model. So are the most capable image and video models (DALL-E, Sora, Imagen, Veo).

Why do companies keep weights closed?

Several reasons:

• Business model — if competitors can copy your weights, the moat disappears. Closed weights protect the multi-hundred-million-dollar training investment.
• Safety controls — the company can update the model, add safety filters, and monitor for misuse without releasing it “into the wild.”
• Compute requirements — even if released, the largest models can’t practically run on consumer hardware. Open weights of a 2-trillion-parameter model would be impractical for most users.
• Liability — if a model is misused, the company has more control over what runs in production when they own the weights.

How does the closed-weight model affect users and developers?

For end users, closed-weight models offer the most polished experience — the company optimizes the model, the UI, and the infrastructure together. ChatGPT works well partly because OpenAI controls every part of the stack.

For developers, closed-weight means:

• Pay per-token API usage (currently typically $1-$30 per million tokens depending on model and provider)
• Subject to provider rate limits, content policies, and service availability
• Data privacy depends on the provider’s policy (most enterprise APIs don’t train on your data)
• Easy to integrate — just an HTTP API call
• No infrastructure to run, scale, or maintain

In 2026 most production AI systems are hybrids: closed-weight APIs (Claude, GPT) for the hardest tasks, open-weight models (Llama, Mistral, DeepSeek) for routine work where cost and privacy matter.

Related terms

• Open Weights
• OpenAI
• Anthropic
• Google DeepMind
• API

Learn more on Beginners in AI

• AI Glossary

Sources and further reading

• OpenAI Platform Documentation
• Anthropic API Documentation

Last reviewed: May 2026. AI terminology evolves quickly — verify specifics on the official source pages above.

You may also like

• AI Glossary
• Open Weights
• OpenAI
• Anthropic
• Google DeepMind