Application Programming Interfaces for AI Models

An Application Programming Interface (API) is a defined set of rules through which one software application communicates with another. For a remotely hosted AI model, the API lets a user send input to the provider's server and receive the model's output without downloading or directly operating the model.

Request and response process

The provider publishes an API endpoint, which is a network address accepting requests in a specified format. Communication commonly occurs through HTTPS using structured data formats such as JSON.

• The client application sends a prompt or other input, model parameters and an authentication credential to the endpoint.
• The remote server validates the request, converts the input into a model-compatible form and runs inference using the hosted model.
• The server returns generated text, predictions or other outputs, along with status information, to the client application.

How the provider controls access

Because computation occurs on the provider's infrastructure, the provider can manage model versions, computing resources and permitted uses while offering the same service to many clients.

• Authentication identifies the client, while authorisation determines which models or functions it may use.
• Rate limits, usage quotas and pricing rules can regulate access and allocate computing capacity.
• Logging and content safeguards may support security and policy enforcement, but they also create privacy and data-governance concerns when user inputs are transmitted to an external server.

API access and open-weight access

An API exposes a model's capabilities as a service, not necessarily its internal parameters, training data or source code. By contrast, an open-weight model makes trained parameters available under stated licence conditions, allowing users to deploy and modify the model more independently.

• API access can simplify deployment because the provider maintains the infrastructure and updates.
• Dependence on a remote provider may create costs, service discontinuity, limited customisation and vendor lock-in.
• An API can serve either a proprietary model or an open-weight model hosted by a third party.