OpenAI released three voice models on May 7, rebuilding its Realtime API line around agents that can listen, reason, speak, and call tools inside one live conversation.
The lineup is GPT-Realtime-2, a conversational model designed to reason while speaking; GPT-Realtime-Translate, which handles real-time interpreting from 70 input languages into 13 output languages; and GPT-Realtime-Whisper, a streaming speech-to-text model. All three are entering the generally available Realtime API, alongside support for remote MCP server calls, image input, and SIP phone dial-in.
Realtime-2 tries to remove the awkward pause
The main change is Realtime-2's ability to keep talking while it reasons through a request. Earlier voice agents often forced users to wait while the model listened, paused, thought, and then answered. In a phone call, a delay of more than a few seconds quickly feels like a broken support line.
Realtime-2 is built to continue the exchange without losing context when a user interrupts, to make parallel tool calls, to use a 128K context window instead of the prior 32K window, and to let developers tune reasoning effort between speed and deeper deliberation. OpenAI describes it as a model that keeps the conversation moving while it reasons, uses tools, and handles corrections or interruptions.
In practice, the pitch is simple: an AI phone agent should no longer sound like a slow customer-service script.
Zillow and Deutsche Telekom are early proof points
OpenAI highlighted two customers for the launch. Zillow, the major U.S. real-estate platform, has been running Realtime-2 on real phone traffic. OpenAI says the deployment improved call success rates and compliance robustness, but did not publish exact percentages. The repeated emphasis on call-success gains is notable because it suggests a large property platform has moved voice agents into mainstream call flows rather than a limited demo.
That matters because the old bottleneck was not basic recognition. Voice agents could hear the words, but often failed when the next turn required juggling constraints: for example, a buyer asking whether a viewing can happen at 3 p.m. and still finish before a child has to be picked up. Zillow's test implies Realtime-2 is getting past more of those multi-variable turns.
Deutsche Telekom, meanwhile, is testing Realtime-Translate for international customer service, where a German-speaking customer and an English-speaking agent could speak through the system without a human interpreter in the middle.
Pricing is simpler for continuous audio
GPT-Realtime-2 is priced at $32 per million audio input tokens, with cached input at $0.40, and $64 per million output tokens. GPT-Realtime-Translate costs $0.034 per minute, while GPT-Realtime-Whisper costs $0.017 per minute.
The minute-based pricing for Translate and Whisper makes budgeting easier for call centers, meeting notes, and other continuous-audio workloads: 1,000 minutes works out to $34 for Translate or $17 for Whisper. Realtime-2 remains token-priced, but the heavily discounted cached input is aimed at applications that reuse a large system prompt in every conversation, such as support agents.
The voice-agent stack is collapsing
A year ago, voice AI was still split across layers. ElevenLabs was strongest in text-to-speech, Deepgram and Whisper focused on speech-to-text, and the first GPT-4o voice experience still struggled with latency and interruption handling. OpenAI is now trying to compress listening, reasoning, and speaking into one reasoning loop.
That changes the competition. The next comparison is less about who transcribes a phone call most accurately and more about who can make an AI handle the next turn like a human on a live call. ElevenLabs still has a voice-cloning advantage, and Anthropic has not yet made a comparable voice API move this year. But if Zillow's early data holds, call-center software and outsourced support vendors will have to explain the shift in their own numbers soon.
Sources: OpenAI has new voice models that reason,CocoLoop, translate, and transcribe as you speak (9to5Mac); OpenAI's New Voice API Models (StartupHub.ai); Advancing voice intelligence with new models in the API (OpenAI official blog)