OpenAI Ships GPT-Realtime-2 With GPT-5-Class Reasoning and Moves the Realtime API to GA
OpenAI shipped three new voice models on May 7 and moved its Realtime API to general availability, ending a beta period that left many production teams watching from the sideline.
The headline model is GPT-Realtime-2, built on GPT-5-class reasoning and designed for live audio-in, audio-out interactions. The context window grows from 32K to 128K tokens — enough for long-session voice agents without external state compression. Developers can tune reasoning intensity across five levels from minimal to xhigh, with default set to low to keep latency predictable.
Two companion models round out the release:
- GPT-Realtime-Translate — live speech translation across 70+ input languages into 13 output languages at $0.034/min
- GPT-Realtime-Whisper — streaming transcription for low-latency speech-to-text at $0.017/min
Benchmark Numbers
OpenAI published two audio-native benchmark comparisons against the prior GPT-Realtime-1.5:
| Benchmark | GPT-Realtime-2 | GPT-Realtime-1.5 | Delta |
|---|---|---|---|
| Big Bench Audio (high) | 96.6% | 81.4% | +15.2 pp |
| Audio MultiChallenge (xhigh) | 48.5% | 34.7% | +13.8 pp |
The Big Bench Audio result is near ceiling on that benchmark. Audio MultiChallenge still shows the model failing more than half the hardest multi-turn dialogue tasks at maximum effort — but a 13.8-point gain is not marginal.
The most concrete production signal is from Zillow: on its internal adversarial benchmark — the call scenarios where prior models broke down — GPT-Realtime-2 achieved a 95% call-success rate, up from 69%. That 26-point improvement on hard cases is a different signal than benchmark performance; it maps to scenarios enterprises have historically had to route to human agents.
Pricing
GPT-Realtime-2 is billed by token: $32/M audio input tokens ($0.40/M for cached input) and $64/M audio output tokens. The cache discount is substantial for deployments with repeated system prompts or knowledge bases. First-audio latency ranges from 1.12s at low reasoning to 2.33s at xhigh.
Architecture Shift
The release separates voice AI into three discrete orchestration primitives rather than forcing everything through a single model. GPT-Realtime-2 handles reasoning and conversation. GPT-Realtime-Translate handles multilingual speech conversion. GPT-Realtime-Whisper handles transcription as a standalone pipe.
The practical effect: enterprises building complex voice stacks can assign workloads to the appropriate model instead of engineering workarounds for a single general-purpose voice model. Mistral’s Voxtral lineup pursues a similar decomposition strategy.
Two new voices — Cedar and Marin — ship alongside the models. All three are available immediately through the OpenAI Realtime API with no waitlist.