Introducing gpt-realtime and Realtime API updates
OpenAI
OpenAI shipped gpt-realtime, a beefed-up speech-to-speech model, plus API updates for phone calls and image input. It's a big step toward voice apps that actually sound and think like a real assistant, not a laggy chatbot with a mic.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI just pushed its Realtime API into a new phase, and the headline is a model called gpt-realtime — a speech-to-speech system built to sound less like a text-to-speech bot bolted onto a chatbot and more like an assistant that actually listens and thinks in audio. That distinction matters more than it sounds. Most voice products still work by transcribing speech, running it through a text model, then synthesizing a reply, which adds latency and strips out tone, emphasis, and timing. gpt-realtime is designed to process and respond in the audio domain directly, which OpenAI says makes conversations feel more natural and responsive.
The API updates around the model are arguably just as important as the model itself. Developers can now hook the Realtime API into MCP servers, OpenAI's protocol for connecting models to external tools and data sources, so a voice agent can pull live information or trigger actions mid-conversation instead of being stuck with whatever it already knows. Image input is also landing, meaning a voice session can now reference something visual — a screenshot, a photo, a diagram — without switching modes or apps.
Then there's SIP support, which is the quieter but genuinely practical addition. SIP is the standard protocol that traditional phone systems and VoIP platforms run on, and by supporting it, OpenAI is opening a direct path for its voice models to plug into call centers, IVR systems, and regular old phone lines. That's a very different market than chat-app voice features — it's the infrastructure businesses actually use to handle customer calls at scale, and it signals OpenAI wants a piece of that world, not just consumer-facing voice assistants.
Taken together, this release reads less like a flashy demo and more like OpenAI quietly filling in the plumbing needed for voice AI to show up in places people don't expect yet: help desks, phone support lines, and tools that mix voice with visual context. The model quality upgrade grabs the headline, but the SIP and MCP additions are the parts that determine whether this actually changes how businesses build with voice, rather than just how good a demo sounds on stage.
My take — AI-written commentary, not fact-checked reporting
The SIP support is the tell here — nobody adds phone-system compatibility for a cool demo, they add it because enterprise sales teams are asking for it. OpenAI is clearly done treating voice as a consumer novelty and is going after the unglamorous, lucrative world of call centers and IVR replacement, which honestly is where the real money in voice AI has always been.
Read more about this at: OpenAI