OpenAI Launches Voice Live with SearchGPT and Real-Time Multimodal API

On September 21, 2026, OpenAI delivered one of the most anticipated upgrades to consumer and developer artificial intelligence: Voice Live inside ChatGPT now connects directly to SearchGPT, enabling fluid, real-time spoken conversations enriched with live web facts. Concurrently, the company opened its Realtime Multimodal API worldwide, streaming simultaneous speech, vision, and text at sub-300ms roundtrip latency. The release mounts immediate pressure on Google’s Gemini Live and Anthropic’s Claude developer stack.

Quick answer: what launched in ChatGPT today?

ChatGPT’s Advanced Voice Mode can now browse the live internet mid-sentence without pausing the conversation, delivering real-time news, scores, and flight data verbally with proper attributions. For software teams, the Realtime Multimodal API replaces fragmented pipelines of speech-to-text, LLM reasoning, and text-to-speech with a unified low-latency model supporting native interruption handling.

How SearchGPT powers live voice conversations

Conversing with voice assistants historically forced a painful tradeoff: either the model delivered instant replies restricted to static training snapshots, or it stalled for multiple seconds while external web retrieval finished. OpenAI resolved this bottleneck by implementing speculative search querying:

• Predictive background search: as the user speaks, semantic embedding classifiers dispatch background queries to SearchGPT before the sentence finishes;
• Natural voice attributions: the model synthesizes conversational citations (“according to Reuters earlier this morning…”) instead of reading aloud awkward website links;
• Barge-in cancellation: low-latency acoustic models allow users to interrupt the assistant naturally, even through speakerphones.

Spec sheet: OpenAI Realtime API vs competitors

Why this matters to you

Conversational computing requires reaction times matching biological human speech, typically between 200 to 300 milliseconds. When delays exceed half a second, human conversational dynamics collapse into awkward turn-taking pauses. By tying sub-second voice synthesis directly to live web indexing and computer vision, OpenAI turns voice assistants from scripted novelty tools into genuine real-time cognitive companions.

For everyday consumers, this capability translates to hands-free multitasking: asking for live flight delay updates, calculating currency conversions while traveling, or diagnosing household plumbing issues by simply holding the phone camera toward a leaking valve while talking. For software engineering teams and enterprise startups, the API unlocks genuinely responsive customer support agents, interactive language tutors, and emergency field inspection assistants without building proprietary telecommunication stacks from scratch.

Frequently asked questions

Who gets access to Voice Live with SearchGPT today?

The update is rolling out starting September 21 to ChatGPT Plus and Team subscribers on iOS and Android, with expansion to free-tier users scheduled over the coming weeks.

Does the Realtime Multimodal API support vision input?

Yes. Developers can stream camera frames alongside microphone audio, allowing agents to comment on what the camera sees in real time with continuous spoken explanations.

What is the pricing model for the Realtime API?

OpenAI charges approximately $0.06 per combined audio minute, accompanied by standard prompt token rates for long conversational context retention.

At DigitalRadar, we track the frontier of artificial intelligence as it happens. Stay on the radar.