14 March 2023

OpenAI releases GPT-4, the first of its models to accept images

OpenAI released GPT-4 on 14 March 2023, a model accepting both image and text input and producing text. Its technical report gives human-level performance on a range of professional and academic examinations, including a simulated bar exam in roughly the top ten per cent of test takers.

Architectures and models Capability threshold Deployed

Background

GPT-3 had been available since 2020 and ChatGPT since November 2022, both text only. Each earlier model in the series had arrived with a paper describing its size, its architecture and the data it was trained on, which is how the field had always worked.

What happened

OpenAI released GPT-4 on 14 March 2023. It accepts image and text input and produces text. The technical report, posted to arXiv the following day, reports human-level performance on a range of professional and academic examinations, a simulated bar exam in roughly the top ten per cent of test takers, and 86.4 per cent on MMLU.

The report also describes something less often quoted: performance was predicted in advance from models trained with a thousandth of the compute. The behaviour of the full system was known before it was built.

Two qualifications belong with the release. An early version had been running in Microsoft’s Bing Chat since February 2023, a month earlier. And image input was available through the API rather than in ChatGPT, so most users met a text-only model.

What followed

The report gave no model size, no architecture and no account of the training data, saying that the competitive landscape and safety implications made disclosure unwise.

That was the change. A laboratory that had published its methods in detail stopped, and the reason given was not that the work was unfinished but that telling anyone was now judged unsafe. Every comparable release since has followed the same practice, and independent assessment of frontier systems has had to proceed without the information that used to accompany them.

Why it mattered

The report withheld the model size, the architecture and the training data, stating that competition and safety made disclosure unwise. That silence from a laboratory whose earlier papers had described their models in detail marked the point at which frontier systems stopped being documented, and every comparable release since has followed it.

Sources

GPT-4 Technical Report. arxiv.org. Primary source
GPT-4. en.wikipedia.org. Secondary

Cite this page

AI Achievements. (2023). OpenAI releases GPT-4, the first of its models to accept images. Retrieved 2026-08-29, from https://achievements.ai/milestone/openai-releases-gpt-4

@misc{achievements_openai_releases_gpt_4,
  title  = {OpenAI releases GPT-4, the first of its models to accept images},
  author = {{AI Achievements}},
  year   = {2023},
  url    = {https://achievements.ai/milestone/openai-releases-gpt-4}
}