OpenAI Ships GPT-Realtime-2 — A Voice Model That Reasons Inside the Audio Loop

On May 7, 2026, OpenAI dropped three new voice models: GPT‑Realtime‑2, GPT‑Realtime‑Translate, and GPT‑Realtime‑Whisper. The flagship brings GPT‑5‑class reasoning running inside the audio loop — not bolted on between transcription and synthesis. TechCrunch reported OpenAI's goal: "move real‑time audio from simple call‑and‑response toward voice interfaces that can actually do work."

The context window jumps from 32K to 128K tokens. Reasoning effort is configurable across five levels. OpenAI's developer docs confirm production features that teams previously built through middleware are now first‑class:

• Preambles: "Let me check that" while calling tools
• Parallel tool calls: Fire multiple backend requests simultaneously
• Recovery behavior: Graceful failure handling without freezing
• Tone control: Deliberate tone adjustment for context

The Next Web noted reasoning happens inside the audio loop — an architectural difference from stitched‑together stacks.

GPT‑Realtime‑Translate handles 70+ input languages and outputs into 13 languages in real time. GPT‑Realtime‑Whisper provides streaming speech‑to‑text. OpenAI's launch blog positions these for live captions, meeting notes, and continuous voice‑agent understanding. Both billed by the minute.

GPT‑Realtime‑2 scored 15.2% higher than Realtime‑1.5 on Big Bench Audio and 13.8% higher on Audio MultiChallenge. The Next Web reported customer results: Zillow saw a 26‑point lift in call‑success rate (69% to 95%). BolnaAI reported 12.5% lower word error rates on Hindi, Tamil, and Telugu.

GPT‑Realtime‑2: $32/1M audio input tokens, $64/1M output, $0.40/1M cached (~$0.048/min raw). Translate at $0.034/min and Whisper at $0.017/min are priced aggressively. Deepgram's Voice Agent API runs $4.50/hr ($0.075/min). ElevenLabs ElevenAgents charges $0.080/min (burst: $0.160/min). A typical multi‑vendor voice stack runs $0.10-$0.25/min. OpenAI's single‑model approach aims below that floor.

Until now, production voice agents ran on a multi‑vendor stack: Whisper/Deepgram for transcription, ElevenLabs/Cartesia for TTS, GPT‑4/Claude for reasoning, plus custom orchestration. The Next Web described GPT‑Realtime‑2 as a direct replacement. Teams optimizing for latency, simplicity, and cost at scale now have a single‑vendor option. TechCrunch noted built‑in safety guardrails halt conversations that violate content guidelines.