What are the main differences between GPT-4 and GPT-3.5?
The differences are mainly evident in complex tasks. GPT-4 is better suited for detailed instructions, creative organization, and analysis with multiple conditions; the gap may not be obvious in simple casual chats. When choosing, compare content accuracy, constraint adherence, and the number of revisions using the same set of real tasks.
Does gpt-4 support image input or image generation?
The current official gpt-4 model documentation lists text for both input and output. GPT-4's early research release introduced multimodal directions, but that does not mean the exact invocation ID on this page supports vision or drawing; when you need to view images, choose models that explicitly list vision capabilities, such as GPT-4o or GPT-4.1, and choose a dedicated image endpoint to generate images.
How can I have GPT-4 continue a previous discussion?
When using Chat Completions, submit the necessary user and assistant history in messages. This can continue the discussion, but it does not imply permanent memory, nor should irrelevant content accumulate indefinitely.
Is gpt-4 the same as the 32K or a fixed-date version?
No. At launch, the official context length for gpt-4 was 8,192 tokens, while gpt-4-32k was another variant; gpt-4-0314 was used to lock the snapshot at that time. A name without a date cannot be directly treated as that snapshot; when reproducing a version, record the full model name and request settings.
Which endpoint should I choose first when calling GPT-4?
Use Chat Completions or Responses and provide the full model ID. Chat Completions uses messages and choices, while Responses uses input and the corresponding response structure; handle history management, streaming events, and tool parameters separately according to the selected interface, and do not mix the two formats.
