GPT-4

14 March 2023: text first

OpenAI announced GPT-4 on 14 March 2023 as a large multimodal model capable of accepting text and image inputs and producing text outputs.

What reached users that day was text. ChatGPT Plus subscribers received capped access, and developers could join an API waitlist, with invitations starting on launch day and scaling gradually. Image input remained a non-public research preview and limited alpha, initially prepared with Be My Eyes.

Context windows and the 25,000-word claim

The standard model carried an 8,192-token context. A 32,768-token version existed with limited access, and it was that longer variant — not the default every Plus subscriber used — behind OpenAI's "more than 25,000 words" description.

Launch API pricing for standard GPT-4 was $0.03 per 1,000 prompt tokens and $0.06 per 1,000 completion tokens.

What OpenAI published, and what it withheld

OpenAI reported six months of alignment work, adversarial testing and safety improvements before deployment, and released its Evals framework as open-source software so outside users could document failures and regressions.

It kept the model's size and architecture undisclosed while publishing a technical report, a system card and pages of benchmark results — including a simulated Uniform Bar Exam score around the top 10% of test takers, which were OpenAI's own controlled evaluations.

Microsoft confirmed the same day that its new Bing preview had already been running an early, search-customized version of GPT-4 for about five weeks.