GPT-4V (GPT-4 with vision)

A GPT-4 that can look at images

GPT-4V — the "V" stands for vision — is a version of GPT-4 that accepts images as input in addition to text. OpenAI documented it in the GPT-4V(ision) system card dated 25 September 2023 ^[1]. In practice it works the way the name suggests: the user gives the model an image together with instructions, and the model analyzes that image input and answers in text ^[5]. The image is not a separate tool that the model calls; it is part of the prompt, handled by the same model that handles the words.

Where it came from

Image input was not a surprise addition. When GPT-4 was launched in March 2023, its ability to accept images was already part of the announcement ^[4]; what arrived with the September system card was the documentation of that capability and its release as something people could use ^[1]. The timing mattered for a second reason: on the same day, OpenAI announced image and voice capabilities for ChatGPT under the title "ChatGPT can now see, hear, and speak" ^[1][2]. The underlying model capability and the consumer-facing product therefore landed together, which is how most people met the feature.

What the system card covers

A system card is the document in which OpenAI sets out what a model can do and how it was evaluated for safety; for GPT-4V it describes the safety properties and evaluations that came with adding image input to GPT-4 ^[1]. That framing points at why the release was treated as more than a feature update. OpenAI describes bringing additional modalities — images among them — into large language models as a key frontier in AI research and development ^[5]. A text-only model is limited to what can be written down; an image input opens tasks that involve looking at a picture, and the model can address them because the picture arrives as part of the prompt.

For developers

Developers reached the model through the API under the names gpt-4-vision-preview and gpt-4-1106-vision-preview ^[3]. The word "preview" in both names signals that these were not treated as final, stable endpoints; both were eventually shut down, on 6 December 2024 ^[3]. Code written against those model names had to be pointed at a successor before that date.

Where it sits in the line

GPT-4V is best understood as a step rather than a destination. It was built on GPT-4 ^[2], and image understanding became native in GPT-4o in May 2024 ^[2]. The trajectory is a familiar one for language models: a capability first appears as an add-on to an existing model, is documented and exposed through an API and a chat product, and is then folded into the next generation, where handling more than one kind of input is part of the design rather than an extension of it.