OpenAI splits realtime voice into three models: reasoning, translation, and transcription

Three models instead of one voice endpoint

OpenAI announced three audio models in the API rather than a single upgraded voice endpoint. GPT‑Realtime‑2 is described as the first voice model with GPT‑5‑class reasoning. GPT‑Realtime‑Translate handles live translation from 70+ input languages into 13 output languages while keeping pace with the speaker. GPT‑Realtime‑Whisper is a streaming speech-to-text model that transcribes as the speaker talks.

The split matters because it maps to how OpenAI frames the work: understanding intent, translating across languages, and producing text are treated as separable functions. A developer building a support flow that only needs live transcription no longer has to route through a full reasoning model, and a translation experience can lean on the dedicated translate model rather than prompting a general one.

Reasoning effort becomes a latency dial

GPT‑Realtime‑2 exposes adjustable reasoning effort across minimal, low, medium, high, and xhigh levels, with low as the default. This is the most operationally significant control in the release: it lets teams trade latency against deliberation on a per-interaction basis, keeping straightforward turns fast while reserving heavier reasoning for complex requests.

OpenAI ties the higher settings to specific eval gains. It reports GPT‑Realtime‑2 (high) scoring 15.2% higher on Big Bench Audio than GPT‑Realtime‑1.5, and GPT‑Realtime‑2 (xhigh) scoring 13.8% higher on Audio MultiChallenge for instruction following. The framing is careful—these are the top settings, and low is the default—so the reasoning wins the benchmarks describe are not what most sessions will use out of the box.

The context window also grows from 32K to 128K, which OpenAI positions for longer, more coherent sessions and multi-step task flows rather than raw model quality.

Designing around the awkward silences of tool use

Several of the new features address a specific failure mode in voice agents: dead air while the system works. Preambles let the model say short phrases like "let me check that" before responding. Parallel tool calls can be made audible with narration such as "checking your calendar." And stronger recovery behavior lets the model say "I'm having trouble with that right now" instead of failing silently.

These are interaction-design decisions surfaced as API capabilities. They acknowledge that a voice agent calling tools mid-conversation needs to signal work-in-progress the way a human would, because silence reads as a broken call rather than a busy assistant.

The three usage patterns OpenAI is betting on

OpenAI names three patterns it sees developers building: voice-to-action, systems-to-voice, and voice-to-voice. Zillow is cited building an assistant that can act on requests like finding homes within a BuyAbility budget and scheduling tours. Deutsche Telekom is building support where customers speak their preferred language while the model translates. Priceline is working toward managing entire trips by voice, including translation once travelers are on the ground.

The systems-to-voice pattern is illustrated by a proactive travel example:

Your inbound flight is delayed, but you can still make your connection. I found the new gate, mapped the fastest route through the terminal, and your bag is still expected to transfer.Montana Labs

That example is doing more than transcription or translation—it turns backend context into unprompted spoken guidance, which is why OpenAI notes the patterns can combine within a single product like Priceline's.

What the split means for teams building voice agents

The specific implication of this release is that OpenAI is asking developers to architect voice apps as compositions of distinct models and tunable settings, not as one call to a black-box assistant. Choosing between Realtime‑2, Translate, and Whisper, setting a reasoning-effort level, and deciding whether to enable preambles and audible tool narration are now explicit design choices that shape latency, cost, and how the agent behaves when a request gets hard.

For applied teams, the practical work moves toward matching each interaction to the right model and reasoning tier—defaulting to low effort, escalating only when a request warrants it, and using the transparency and recovery features to keep conversations intact when tools are slow or fail.