On May 7, 2026, OpenAI introduced three new audio models in the Realtime API: GPT‑Realtime‑2, GPT‑Realtime‑Translate, and GPT‑Realtime‑Whisper. The launch matters because it gives developers a new voice stack for reasoning, live translation, and streaming transcription at the same time that OpenAI’s older Realtime API beta and several legacy preview audio models hit their shutdown date.
That combination makes this more than a normal product update. For teams building voice agents, support systems, multilingual assistants, and speech-driven workflows, OpenAI has turned May 7 into a real migration point as well as a feature launch.
What OpenAI shipped on May 7
OpenAI framed the release as a new generation of realtime voice models for the API. The three launches target different layers of spoken interaction:
- GPT‑Realtime‑2 is the new flagship voice model for live conversations that need reasoning, tool use, and better conversational recovery.
- GPT‑Realtime‑Translate is a live speech translation model that supports more than 70 input languages and 13 output languages.
- GPT‑Realtime‑Whisper is a streaming speech-to-text model designed for low-latency transcription.
OpenAI said GPT‑Realtime‑2 adds several features that matter for production systems rather than voice demos: short audible preambles while the model is working, parallel tool calls, stronger failure recovery, more controllable tone, and adjustable reasoning effort settings ranging from minimal to xhigh. It also expands the context window for these sessions from 32K to 128K, which is a meaningful jump for longer conversations and more complex task flows.
Pricing is also explicit from day one. OpenAI priced GPT‑Realtime‑2 at $32 per 1 million audio input tokens, $0.40 for cached audio input tokens, and $64 per 1 million audio output tokens. GPT‑Realtime‑Translate is priced at $0.034 per minute, while GPT‑Realtime‑Whisper is priced at $0.017 per minute.
The timing is especially important. OpenAI’s deprecations documentation shows the Realtime API beta is removed on May 7, 2026, and older GPT‑4o realtime and audio preview models are also shut down the same day. That means builders are not just evaluating a new option. Many of them are being pushed onto a new default path.
Why GPT‑Realtime‑2 is bigger than a typical voice model refresh
Plenty of voice products can already transcribe speech or speak back an answer. The harder problem is building a voice agent that can stay useful while the task becomes messy: the user interrupts, changes direction, asks for something more complicated, or needs the system to call tools without breaking the rhythm of the conversation.
That is the gap OpenAI is trying to close with GPT‑Realtime‑2. The core product story is not just faster speech. It is live reasoning plus action. OpenAI is packaging voice interaction, tool use, context retention, and recovery behavior into one runtime layer that is much closer to how production agents actually operate.
This is also why the extra features matter. Preambles help the system feel responsive while work is happening in the background. Parallel tool calls reduce waiting. Better recovery behavior matters because dead air and brittle failures are much more damaging in a spoken workflow than in a text box. A 128K context window matters because real support and operations calls do not stay short for long.
OpenAI’s own examples point to that production angle. The company highlighted early work from Zillow, Priceline, Deutsche Telekom, Vimeo, and others. Zillow in particular said GPT‑Realtime‑2 delivered a 26-point lift in call success rate on its hardest adversarial benchmark after prompt optimization, which is the kind of operational metric enterprises will care about more than a demo clip.
Where the business impact will show up first
Support and contact-center workflows
Customer support is the clearest near-term fit. A voice agent that can acknowledge work in progress, query tools, recover from errors, and keep talking naturally is much more useful than a basic IVR replacement. This release pushes closer to that standard.
Multilingual service operations
Live translation is the second big signal. Many companies have handled multilingual support by stitching together separate transcription, translation, and response systems. OpenAI is now offering a more unified path inside the same realtime stack. That could be attractive for support teams, travel operations, global sales, events, and education workflows where speed matters as much as fluency.
Speech-first workflow automation
Streaming transcription is not glamorous on its own, but it is operationally important. Low-latency transcription can power live captions, meeting notes, field-service documentation, healthcare intake, recruiting calls, and downstream automations that should start before the conversation ends. GPT‑Realtime‑Whisper makes that layer easier to productize inside a broader voice workflow.
OpenAI also said the Realtime API supports EU Data Residency for EU-based applications and remains covered by its enterprise privacy commitments. That matters because voice systems often get blocked less by raw model quality than by compliance, data handling, and operational trust.
What to watch next
The next question is not whether voice AI exists. That question is settled. The real question is whether companies will now consolidate separate speech, translation, and agent orchestration layers into fewer services as these model platforms get more capable.
OpenAI is clearly trying to become that fuller voice runtime, not just a speech model vendor. The same-day removal of older realtime paths strengthens that message. Builders now have a clearer product line, a clearer migration path, and a clearer signal about where OpenAI wants live voice development to go next.
For AI agents and enterprise automation, the practical implication is straightforward: voice is no longer just a user interface on top of an agent. It is becoming part of the agent runtime itself. Teams automating support, intake, scheduling, field operations, or multilingual service should treat this release as a concrete reason to re-evaluate where spoken interaction belongs in their workflow stack.