Manage models for GPT for Word

Models are the engines that power GPT for Word. GPT for Word supports models from OpenAI, Perplexity, Google, Anthropic, Parallel, DeepSeek, Mistral, and xAI. GPT for Word also supports Azure models and open-source models through any OpenAI-compatible API endpoint.

info

As a space owner or admin, you can control which models are available to users in your space through the AI configuration section in the GPT for Work dashboard.

Which model is the right one depends on your specific needs in terms of response quality, speed, privacy, and cost. The following table summarizes the kinds of models available in GPT for Word, from zero setup to more advanced options:

Use	Setup	When
Built-in models	No setup. Use the default model or select a different model.	You want to get started quickly without setting up API keys. ✅ Available models
Provider-connected models 🔑	Set an API key for an AI provider.	You want to benefit from the terms of service you have with an AI provider, such as cost, privacy, and rate limits. You want to use a model that is not available as a built-in model. You want more control. ✅ Available models
Endpoint-connected models from cloud platforms	Azure OpenAI models Set up an Azure OpenAI API endpoint.	Your company is already using Azure with a consumption commitment. Using Azure OpenAI counts towards that commitment. You want to use models that are available through Azure OpenAI. ✅ Available models
Endpoint-connected models from cloud platforms	OpenAI-compatible models Set up an OpenAI-compatible platform endpoint.	You want to use models available on a cloud-based AI platform such as Anyscale or Fireworks AI. You need a fast and powerful inference infrastructure that scales to your performance needs.
Endpoint-connected models from local servers	OpenAI-compatible models Set up an OpenAI-compatible server endpoint.	You want to use models available on a local LLM server such as LM Studio or Ollama. You want to run models on a local machine. You do not want to send or receive any prompt or response data over the internet.

What's next