OpenAI drops three new realtime voice models

Talking to an AI voice agent has always felt uncomfortably brittle. You ask a slightly complex question, get hit with dead air, and wonder whether the bot is thinking, crashing, or quietly hanging up on you.

OpenAI just launched a trio of new real-time audio models in its API to fix that disconnect. Leading the pack is GPT-Realtime-2, which OpenAI bills as its first voice model with GPT-5-class reasoning. Alongside it come GPT-Realtime-Translate for live multilingual voice translation and GPT-Realtime-Whisper for low-latency streaming transcription.

Talking while thinking

The biggest leap in GPT-Realtime-2 is that it bridges the gap between natural speech and heavy reasoning without freezing the conversation.

The model introduces preambles and tool transparency. When fetching data or checking systems, it naturally murmurs phrases like "let me check that" or "looking that up now." It can execute multiple parallel tool calls while audibly keeping the user in the loop.

Under the hood, the context window jumps from 32K to 128K, making room for long, context-heavy agent workflows. Developers can also toggle reasoning effort across five tiers, from minimal up to xhigh, with low set as the default balance between latency and depth.

On the Big Bench Audio benchmark, high-tier GPT-Realtime-2 scored 96.6%, jumping 15.2 points over its predecessor. Instruction-following tests also saw a 13.8% boost, showing better grip on specialized industry terms and graceful recovery when interrupted.

Live translation and streaming text

To round out a complete voice agent pipeline, OpenAI also shipped two specialized live models.

GPT-Realtime-Translate handles speech input from over 70 languages and converts it on the fly into 13 output languages while matching the speaker's cadence. Deutsche Telekom is testing it for multilingual call centers, while Vimeo demonstrated it delivering instant live audio dubbing for tutorial videos.

GPT-Realtime-Whisper covers live streaming speech-to-text for real-time captioning and running meeting summaries. Pricing reflects the split: GPT-Realtime-2 runs on tokens ($32 per million audio input tokens, $0.40 cached, and $64 per million audio output tokens), while Translate and Whisper run by the clock at $0.034 and $0.017 per minute, respectively.

Why it matters

AI labs spent years obsessing over text-based autonomous agents, but the real enterprise prize will be spoken, not typed.

Until now, building a dependable voice assistant required stitching together transcription, LLM reasoning, orchestration code, and speech synthesis. That patchwork usually produced awkward two-second delays. With Priceline and Zillow already testing end-to-end voice flows, complex voice commerce is finally becoming viable.

We think OpenAI's decision to prioritize conversation management over vocal parlor tricks is spot-on. Giving the model the ability to talk while it executes tools solves the psychological barrier of dead air in voice interactions.

That said, high-volume production remains a financial minefield. Pure audio tokens are substantially pricier than text tokens, and while active safety classifiers can intercept rogue conversations, operating live voice agents with full tool access still presents massive customer support safety and compliance hurdles.