OpenAI has added three new voice models to Realtime API, and the pitch is pretty clear: stop treating voice as a demo feature and start treating it like infrastructure. The flagship GPT-Realtime-2 is the big one, with GPT-5-level reasoning, a much larger context window, and better handling of pauses, interruptions, and topic switches. OpenAI is positioning the new voice models in Realtime API against Perplexity, Google, Zoom, and a crowd of startups that all promise ”natural” voice until the conversation gets messy.
The company is also betting that reliability, not a charming synthetic accent, is what will sell. In customer support, booking flows, and enterprise tools, a voice agent that keeps its place and calls tools correctly is more useful than one that simply sounds pleasant for 20 seconds.
GPT-Realtime-2 gets the main upgrade
GPT-Realtime-2 is OpenAI’s first voice model with reasoning on the level of GPT-5. The context window has expanded from 32K to 128K, and developers can tune reasoning effort from ”minimal” to ”xhigh”. OpenAI says the model is better at dealing with interruptions, mistakes, and abrupt topic changes, which is exactly where many voice assistants fall apart and start sounding like they have lost the plot.
The benchmark gains are solid rather than flashy. GPT-Realtime-2 (high) is 15.2% better than GPT-Realtime-1.5 on Big Bench Audio, while the xhigh version is 13.8% ahead on Audio MultiChallenge. That matters because enterprise buyers care less about novelty and more about whether the system can survive a real conversation without derailing a call or inventing confidence where none exists.
”What impressed us about GPT-Realtime-2 was the intelligence and reliability of tool calling in complex voice scenarios. On our toughest adversarial benchmark, that translated into a 26-point increase in successful calls after prompt optimization, 95% versus 69%. The model is also noticeably stronger on Fair Housing compliance, which is critical for our business.”
Josh Weisberg, senior vice president and head of AI, Zillow
Realtime Translate and Whisper target practical use cases
The other two models are less dramatic on paper but arguably easier to monetize. GPT-Realtime-Translate supports more than 70 input languages and 13 output languages for live translation, while GPT-Realtime-Whisper provides low-latency streaming speech transcription. One is for conversation across languages; the other is for subtitles, notes, and any workflow where waiting for the speaker to finish is already too slow.
That puts OpenAI directly into territory long occupied by Zoom, Google Meet, and a parade of enterprise transcription vendors. The differentiator is supposed to be robustness with regional accents and domain-specific terminology, which is the part many polished demos quietly skip over. Deutsche Telekom is testing multilingual voice communication, Vimeo is showing real-time translation for training video, and BolnaAI says it is seeing lower speech-recognition error rates for Hindi, Tamil, and Telugu.
Pricing for the three voice models
OpenAI has also put prices on the table. GPT-Realtime-2 costs $32 per 1 million audio input tokens and $64 per 1 million audio output tokens. GPT-Realtime-Translate is priced at $0.034 per minute, and GPT-Realtime-Whisper costs $0.017 per minute. All three models are available now in Realtime API.
- GPT-Realtime-2: $32 per 1 million audio input tokens
- GPT-Realtime-2: $64 per 1 million audio output tokens
- GPT-Realtime-Translate: $0.034 per minute
- GPT-Realtime-Whisper: $0.017 per minute
OpenAI is also leaning on safety controls, with active session classifiers in Realtime API and custom rules available through Agents SDK. That is the unglamorous part of voice AI, but it is also the part that decides whether these systems are useful in regulated or customer-facing settings instead of becoming a moderation headache with a nice voice.
The next fight is straightforward: which vendors can turn voice from a neat interface into a dependable business workflow? OpenAI is clearly aiming at booking, support, media, and multilingual sales. If rivals want to keep that traffic, they will need more than better pronunciation and a slick demo reel.