GPT-4 — Going Multimodal

GPT-4 is the large language model OpenAI announced on 14 March 2023. From that same day, subscribers to the paid ChatGPT Plus service could use GPT-4, subject to a usage cap. The developer API started by inviting people from a waitlist and opened to every API developer with a history of successful payments on 6 July 2023. The announcement did not restrict availability to particular countries; access came through online services, ChatGPT Plus and the API.

OpenAI introduced GPT-4 as a large multimodal model that accepts image and text inputs and emits text outputs. Image input, however, was not open to the public at launch.

Timeline of the release

Terms on launch day

OpenAI's March 2023 announcement page, which can be checked in the Internet Archive's copy, set out the terms in detail.

On images, the page said: "Image inputs are still a research preview and not publicly available." To prepare the capability for wider availability, OpenAI was working with a single partner, linking to Be My Eyes, an app that assists blind and low-vision people. Image input reached ChatGPT users six months later: on 25 September 2023 OpenAI announced it was rolling out voice and image capabilities to Plus and Enterprise users over the following two weeks.

What was new

OpenAI described GPT-4 as less capable than humans in many real-world scenarios but showing human-level performance on various professional and academic benchmarks. Its example was a simulated bar exam, where GPT-4 scored around the top 10% of test takers; GPT-3.5 had scored around the bottom 10%.

It also described how the model was built. Over two years OpenAI rebuilt its deep-learning stack and, together with Azure, co-designed a supercomputer from the ground up. GPT-3.5 had been trained a year earlier as a first test run of that system, and the GPT-4 training run became OpenAI's first large model whose performance it could predict accurately in advance. The company said it had spent six months iteratively aligning GPT-4 using lessons from adversarial testing and ChatGPT. It also open-sourced OpenAI Evals, its framework for evaluating models.

Exam scores, and a second look

The technical report tabulated results on dozens of exams. A selection:

The most widely quoted figure, the bar-exam percentile, was later questioned. In a 2024 paper in Artificial Intelligence and Law, Eric Martínez showed that the top-10% estimate rested on comparisons with February test-takers, who are heavily skewed towards repeat takers who failed in July and score well below the general population. Using a July sitting, he estimated GPT-4 below the 69th percentile (about the 48th on essays); against first-time takers, about the 62nd (42nd on essays); and against those who passed, about the 48th (15th on essays).

The AP results were uneven too. GPT-4 scored the top grade of 5 in Biology, Environmental Science and Macroeconomics, but a 2 out of 5 in both English exams — the same as GPT-3.5.

What was not disclosed

The technical report described GPT-4 as a Transformer-style model pre-trained to predict the next token in a document, using publicly available data and data licensed from third parties, then fine-tuned with reinforcement learning from human feedback (RLHF). Beyond that, it said nothing:

"Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar."

OpenAI has not published the model's size — its parameter count. That is a sharp contrast with GPT-3, whose paper gave 175 billion parameters. The report also noted that the vast majority of GPT-4's pre-training data cut off in September 2021.

What else was happening

On the day GPT-4 was announced, Anthropic began offering its Claude chat model. Eight days later, on 22 March, the Future of Life Institute published an open letter calling on all AI labs to pause for at least six months the training of AI systems more powerful than GPT-4. GPT-4 had become both a performance benchmark and a reference point for the debate over regulation and safety.

On 13 May 2024 OpenAI announced GPT-4o, which reasons across audio, vision and text in real time. According to Ars Technica, OpenAI's ChatGPT changelog of 10 April 2025 announced that GPT-4 would be retired from ChatGPT and fully replaced by GPT-4o effective 30 April, while remaining available in the API.

Where GPT-4 fits

GPT-4 arrived three and a half months after ChatGPT. OpenAI's computing ran on Microsoft's Azure, a partnership that began with the 2019 investment and had just been extended in January 2023. For the terms see large language model and foundation model; for the wider story, the AI timeline; for the company, OpenAI.