Introducing GPT-4o

OpenAI is rolling out GPT-4o as a new flagship model designed to make advanced AI feel more natural in real time—across voice, text, and vision—while also pushing that capability to free users. The centerpiece is a shift away from the multi-model “voice mode” stack that previously stitched together transcription, reasoning, and text-to-speech with added latency. GPT-4o instead handles those modalities natively, enabling faster, more interruptible conversations and a more seamless back-and-forth experience.

The rollout matters because it targets two friction points at once: accessibility and interaction quality. OpenAI says GPT-4o brings “GPT-4 intelligence” to everyone, including free users, and will be deployed across ChatGPT over the next few weeks. Alongside the model launch, OpenAI is releasing a desktop version of ChatGPT, refreshing the UI to keep users focused on collaboration rather than interface mechanics. The company also highlights earlier moves to reduce signup friction, including making ChatGPT available without a sign-up flow.

GPT-4o’s performance claims are framed around speed, cost, and capacity. OpenAI says GPT-4o is faster, 50% cheaper, and offers five times higher rate limits than GPT-4 Turbo in the API. For paid users, it also claims up to five times the capacity limits compared with free users. Multilingual quality is another emphasis: OpenAI says GPT-4o improves quality and speed across 50 languages.

ChatGPT’s feature set is expanded to match the new model’s multimodal abilities. Users can access GPTs through the GPT Store, where custom ChatGPTs created by builders are available to a larger audience. Vision support is positioned as a core workflow upgrade: users can upload screenshots, photos, and documents containing both text and images, then ask questions about that content. Memory is also introduced as a continuity layer across conversations. Additional tools mentioned include browsing for real-time information and advanced data analysis for working with uploaded charts.

Live demos during the announcement showcased the “real-time conversational speech” capability. In one segment, Mark Chen demonstrates a voice interaction where the model responds with low lag, can be interrupted, and detects cues like fast breathing—then adjusts tone accordingly. The demo also highlights expressive voice generation, including different drama levels and a “robotic” singing-style delivery for a bedtime story.

Other demonstrations focused on vision and multimodal reasoning. Barrett Zoph works through a linear equation using a handwritten equation on paper, with the model providing step-by-step hints rather than just the final answer. A coding-and-plot example shows GPT-4o interpreting a screen, summarizing what’s displayed, and answering questions about the graph’s axes and temperature trends. Audience-request prompts then test real-time translation (Italian↔English), emotion inference from a selfie, and conversational responsiveness.

Finally, OpenAI flags safety as a central challenge for GPT-4o because it involves real-time audio and vision. The company says it is working with stakeholders across government, media, entertainment, and civil society to mitigate misuse as capabilities roll out.