OpenAI introduces GPT-4 with text access and an API waitlist

OpenAI introduced GPT-4 on March 14, 2023. The model accepted text and image inputs and produced text, extending the kinds of material a language model could interpret.

Text access at launch

The initial release made text capabilities available through ChatGPT and an application programming interface (API) waitlist. Image input was not a general public API feature on launch day; OpenAI began with a partner deployment. The distinction mattered for applications built around photographs, charts, or scanned documents.

Evaluations alongside the model

OpenAI also released OpenAI Evals, a framework for testing model behavior and contributing evaluation cases. The announcement reported improvements over GPT-3.5 in reliability and handling complex instructions, while retaining problems such as invented facts and reasoning errors.

GPT-4 expanded the model options available for chat and application development. Choosing it still involved evaluating its answers on the task being automated, rather than relying on a benchmark result alone.