On 7 May 2026, OpenAI announced GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper. The three-model family separated agentic speech, continuous translation and streaming transcription while keeping each workload inside the Realtime API.
This release brief was checked against first-party material on 10 August 2026. The date above is the public announcement date, not the date a repository was created or a third-party provider added the model. Where access or weights arrived later, that distinction is recorded below.
Release record
| Field | Verified detail |
|---|---|
| Announcement | 7 May 2026 |
| Availability or weight release | 7 May 2026 through the OpenAI Realtime API |
| Release type | hosted realtime speech-to-speech, translation and transcription model family |
| Access | OpenAI API access; no downloadable weights |
| Architecture | native realtime audio models, with GPT-Realtime-2 adding GPT-5-class reasoning and tool use |
| Maximum stated context | 128K tokens for GPT-Realtime-2, increased from 32K in the previous generation |
What changed
This was a major model launch rather than a product-only voice update. GPT-Realtime-2 added parallel tool calls, adjustable reasoning effort and a fourfold context increase; Translate covered more than 70 input languages; and Whisper moved transcription into a low-latency streaming model. The release therefore changed both model capability and the architecture available to voice-agent developers.
The practical comparison is therefore not simply whether GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper has the largest headline score. Teams need to ask whether its architecture, access terms, latency, tool behaviour and evaluation setup match the workload they actually intend to run. A model can lead one harness while losing on cost, refusal behaviour, multilingual quality or repeatability in another.
Benchmarks worth retaining
| Evaluation | Reported result | How to read it |
|---|---|---|
| Big Bench Audio | 15.2% higher for GPT-Realtime-2 high than GPT-Realtime-1.5 | OpenAI-reported relative gain on audio intelligence |
| Audio MultiChallenge | 13.8% higher for GPT-Realtime-2 xhigh | Instruction-following gain at the disclosed reasoning setting |
| Zillow call success | 95% versus 69% after prompt optimisation | Early-customer result, not a controlled general benchmark |
These are release-time results, not independently reproduced guarantees. The benchmark gains use named third-party audio evaluations and selected reasoning settings; latency, telephony conditions, tool schemas and language mix can materially change production results. Scores should remain attached to the disclosed effort setting, agent harness, tool access, timeout, context-management policy and judge model. Moving a number into a procurement sheet without those conditions creates false comparability.
Architecture and access
GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper is described as native realtime audio models, with GPT-Realtime-2 adding GPT-5-class reasoning and tool use with 128K tokens for GPT-Realtime-2, increased from 32K in the previous generation of stated context. Its access position at verification time is OpenAI API access; no downloadable weights. That wording matters: open weights, source-available weights, an API, a product preview and a research demonstration give adopters very different rights and different levels of reproducibility.
Before deployment, record the exact model identifier or checkpoint, inference stack, quantisation, reasoning setting, region, price schedule and supplier terms. If the release uses a custom licence, read the licence itself rather than relying on the word “open” in launch copy. If it is API-only, preserve the dated documentation and change-notice route because the served snapshot can change without a downloadable artefact.
What an evaluation should test next
For GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper, a credible internal gate should include:
- a frozen set of representative tasks with pass, fail and abstain criteria;
- a matched baseline using the same tools, timeout, prompt budget and reviewer rubric;
- repeated runs to expose variance rather than reporting a single best attempt;
- latency, token use and total task cost alongside task success;
- adversarial, multilingual and long-context cases relevant to the real deployment; and
- rollback evidence showing the previous model can be restored safely.
The wider model change-control guide explains how to keep model, prompt, tool and corpus changes reconstructable. The AI dependency inventory guide covers the release and supplier records needed after deployment.
AIEngine verdict
The family belongs in the archive because it introduced three named API models with distinct operating roles. Teams should evaluate audio quality, interruption handling, tool reliability and total per-minute cost together.
This is a launch assessment, not a certification. Benchmark leadership is useful evidence of where to test; it is not authorization to place the model in a high-impact workflow without domain evaluation, security review and an accountable owner.
Primary sources
- OpenAI voice-model family announcement
- OpenAI Realtime API guide
- OpenAI audio and voice model documentation
Image provenance
Hero image: OpenAI official release artwork. The locally served WebP is a crop of the first-party release or model-card asset recorded in the repository provenance manifest.
