OpenAI Unveils GPT-Realtime-2 Series with GPT-5-level Reasoning

OpenAI has added 3 new audio models to the Realtime API.

• GPT-Realtime-2 is the first voice model with GPT-5-level reasoning, designed for tool calling, interruption handling, and maintaining long conversations.
• GPT-Realtime-Translate translates 70+ input languages into 13 output languages in real time.
• GPT-Realtime-Whisper is a streaming transcription model that converts speech to text as you speak.

The core change is that voice has expanded beyond simple call-and-response into an interface that listens, reasons, translates, transcribes, and acts. The presented usage patterns include Voice-to-Action, Systems-to-Voice, and Voice-to-Voice.

GPT-Realtime-2 uses Preambles to indicate it's processing, supports parallel tool calling and tool transparency, and recovers naturally without cutting off the conversation even on failure. The context window has grown from 32K to 128K, and tone control and domain terminology understanding have also been strengthened.

On performance, GPT-Realtime-2 (high) improved 15.2% over GPT-Realtime-1.5 on Big Bench Audio, and xhigh improved 13.8% on Audio MultiChallenge. Zillow reported that call success rate rose to 95% vs. 69%, and BolnaAI said its translation WER is 12.5% lower than other models.

Pricing: GPT-Realtime-2 costs $32 per 1M input tokens, $0.40 for cached input tokens, and $64 per 1M output tokens. GPT-Realtime-Translate costs $0.034 per minute, and GPT-Realtime-Whisper costs $0.017 per minute; all three models are immediately available in the Realtime API and Playground.

On the safety side, active classifiers and multi-layered mitigations are applied, and sessions can be terminated if harmful content violations are detected. Developers can attach additional guardrails using the Agents SDK, and EU data residency and enterprise privacy commitments are also supported.