OpenAI launches GPT-4o, a single model for text, audio, and images
OpenAI introduced GPT-4o on 13 May 2024, one model trained across text, image, and audio, demonstrated by chief technology officer Mira Murati in a livestream showing real-time voice conversation. OpenAI gave it to free ChatGPT users and said it handled around 50 languages.
Why it mattered Putting a GPT-4-class model and spoken conversation behind ChatGPT’s free tier moved the technology from a paid professional tool toward an ordinary consumer product.
Two years after ChatGPT opened to anyone with a browser, the strongest model behind it still sat on the paid side of the product. That changed on 13 May 2024. OpenAI presented GPT-4o at a livestreamed event it called the Spring Update and said the model would go to free users of ChatGPT, with higher usage limits kept for subscribers.
Mira Murati, then chief technology officer, ran the demonstration. Two things were shown that the product had not done in public before: a spoken conversation held at close to the speed of human talk, and the model reasoning about what a phone camera was pointed at while that conversation continued. OpenAI said GPT-4o worked in around 50 languages.
The claim underneath the demonstration was architectural. OpenAI described GPT-4o as one model trained end to end across text, image, and audio rather than a text model with speech and vision attached at the edges. The letter stood for omni. A single network took in sound and pictures and returned text or speech.
The release landed in a crowded quarter. Claude 3 Opus had taken the top of several public benchmark tables ten weeks earlier, and Anthropic, Google, and Meta were all shipping model families on a schedule. What GPT-4o competed on was reach. A free tier meant the number of people who could hold a conversation with a frontier model was no longer bounded by who would pay for one.
Four weeks later Apple said its own system-level AI would hand difficult questions to ChatGPT with the user’s permission, making OpenAI’s model the first from outside the company to ship inside Apple’s operating systems. In September OpenAI moved the contest again with o1, a model that works through a problem before it answers rather than replying at once.
Corrections
- 2026-09-01 - This entry originally said GPT-4o handled “more than 50 languages.” OpenAI’s own claim, as quoted by TechCrunch, is “enhanced performance in around 50 languages”; corrected.