What Can GPT‑4 Do? Vision, Long Memory, Safer AI and More

GPT‑4 Release and Access

OpenAI launched GPT‑4, the latest large language model. It is available to ChatGPT+ subscribers and via an API wait‑list.

Multimodal Vision Capability

GPT‑4 can accept image inputs and generate captions, classifications, and contextual analyses. Example: when shown a photo of an iPhone with a VGA connector plugged into its charging port, the model identified the objects and explained the humor of using an outdated connector with a modern device. The image mode is still in research and not publicly open; OpenAI partners with the “Be My Eyes” app to provide a virtual‑volunteer assistance mode for low‑vision users.

Extended Context Length

The model supports up to 25 000 tokens (≈50 pages) of input for long‑form generation and retains a working memory of about 64 000 tokens (≈100 pages) across a conversation, far exceeding the 8 000‑token limit of GPT‑3.5‑based ChatGPT.

Improved Reasoning and Exam Performance

In simulated bar‑exam questions GPT‑4 scored in the top 10 % of test‑takers, outperforming 90 % of human examinees, while GPT‑3.5 scored in the bottom 10 %. The model also shows stronger performance on creative and technical writing tasks and follows finer‑grained instructions more reliably.

Socratic Prompting

Developers can set a system message that instructs the model to adopt a Socratic tutor style, asking guiding questions instead of providing direct answers. Example prompt:

You are a tutor who always responds in a Socratic manner. You never give the student the answer directly; instead you ask probing questions that lead the student to discover the solution themselves.

Safety and Alignment Improvements

OpenAI spent over six months incorporating additional human feedback, hiring more than 50 safety experts, and applying new reinforcement‑learning techniques. Internal adversarial‑testing scores improved by roughly 40 % compared with GPT‑3.5. The model is claimed to be safer and more consistent, though it still lacks verification for malicious code generation.

Remaining Limitations

GPT‑4 continues to exhibit social bias, occasional factual inaccuracies, and its generated code cannot be guaranteed safe.

New Use Cases and Partnerships

Multilingual chat between Spanish and French speakers.

“Explain My Answer” feature that breaks down model responses.

Khan Academy’s “Khanmigo” tutoring assistant powered by GPT‑4.

Virtual‑volunteer assistance through the Be My Eyes partnership.

Availability and Capacity

ChatGPT+ users can access GPT‑4 with usage caps that will be adjusted as capacity grows. Developers can join the API wait‑list through OpenAI’s sign‑up page.