Updated October 15, 2026: This article revisits OpenAI’s September 2023 announcement and separates that launch from current product documentation. The update is based on OpenAI’s published sources, not hands-on testing. Availability and limits can change.

Cover image: The retained artwork’s “Now” headline refers to the 2023 announcement, not a new release or a hands-on test.

On September 25, 2023, OpenAI announced voice conversations and image input for ChatGPT.

What OpenAI announced in September 2023

The planned two-week rollout covered Plus and Enterprise. Voice initially required opting in on iOS or Android; image input covered all platforms. Five voices were created with professional voice actors.

The original voice system used Whisper for transcription and a text-to-speech model for spoken replies. OpenAI named multimodal GPT-3.5 and GPT-4 for image understanding. Those details describe the 2023 launch.

The launch changed the interaction: a question could be spoken, and a picture could provide context that would otherwise need a written description. That is the useful historical distinction. This announcement alone does not establish that ChatGPT was more accurate or capable than competing assistants.

What current users should check

Voice access and limits

The current ChatGPT Voice guide describes web, iOS, and Android access. Free access is limited; paid allowances differ. Available Voice options depend on factors including plan, workspace settings, region, and app version. Check the controls and limit notices in your account.

The guide distinguishes Live, Advanced, and Standard. Live supports text and available image attachments in the same chat. Video and screen sharing are separate capabilities available to eligible mobile subscribers using Advanced; Live does not support them. Voice transcripts may not match the conversation exactly.

Image input and limitations

OpenAI’s Image Inputs FAQ lists Free and paid plan access, subject to account settings and usage limits, on web and mobile. Supported still-image formats include PNG, JPEG, and non-animated GIF, with a 20 MB limit per image.

A clear image and a focused question help define the task. The FAQ warns about incorrect descriptions, approximate object counts, rotated text, and difficulty with precise spatial positions. Enlarge small text without removing needed context. A confident description should still be checked against the original image.

A low-stakes example

This is a fictional illustration, not a product test. Imagine someone planning a small paper-craft display. They have a photo of colored paper, ribbon, and blank cards, with no private information visible. A useful starting prompt would be:

“Suggest two tabletop display ideas using the materials visible in this photo. List anything you cannot identify clearly. Ask me about quantities before giving me a shopping list.”

If voice is available, a follow-up could be: “Talk me through the simpler idea in three short steps. Pause after each step so I can ask a question.”

The person would then count the supplies themselves and compare the proposed steps with what is actually on the table. This keeps the task easy to verify. It also separates a creative suggestion from an observation that needs confirmation.

Review privacy settings before sharing

OpenAI’s Data Controls documentation explains the “Improve the model for everyone” setting. Turning it off prevents new conversations from being used for model training, while leaving them in chat history. If you submit feedback, the associated conversation may still be used for training. Managed workspaces have their own controls; Business, Enterprise, Edu, and Healthcare content is not used for training by default.

The Voice guide separately explains that audio and video clips are not used for training unless you choose to share them. Transcripts and other conversation files may be used depending on your plan and settings. Clip-sharing choices and general model-improvement settings should both be reviewed.

For an everyday visual question, consider whether a cropped or anonymized picture would be enough. For a spoken question, avoid including information about another person that is unnecessary for the task. Choose the smallest amount of material that lets you ask the question clearly.

How to read the original launch today

The September 2023 announcement is a historical reference. Use current documentation and your account’s available controls when deciding what you can do now. Ask for clarification when a reply seems wrong, and verify the details that matter. An assistant revising an answer after a correction is not, by itself, evidence that its underlying model has learned from that exchange.

Image input is a different task from creating an image. For the separate historical launch, read our DALL·E 3 launch retrospective.