OpenAI's New Voice Models Explained — gpt-realtime-2, Realtime Translate, and Realtime Whisper

OpenAI released three new voice models in May 2026, and the announcement is worth paying attention to — not because it changes what voice AI can do today, but because it signals clearly where the category is heading.

The three models — gpt-realtime-2, Realtime Translate, and Realtime Whisper — each address a specific layer of the voice AI stack. Here is what they actually do, what they cost, and what they mean for anyone building voice applications.

The Three New Models

gpt-realtime-2: Voice Intelligence That Can Actually Reason

The most significant release is gpt-realtime-2. Its predecessor, gpt-realtime-1.5, handled straightforward voice conversations competently but struggled with requests that required multi-step reasoning, ambiguous instructions, or tool calls mid-conversation. gpt-realtime-2 closes that gap.

The key claim: gpt-realtime-2 is built on GPT-5-class reasoning. In practice, that means it can handle the kinds of requests that previously required routing to a text-based model:

• Multi-step instructions with dependencies ("book the appointment, and if the 3pm slot is gone, try 4pm, and if not, send a calendar invite for tomorrow morning")
• Ambiguous questions that require clarification-gathering before answering
• Complex workflows with parallel tool calls — the model can trigger multiple integrations simultaneously rather than sequentially
• Natural conversational preambles — short phrases before the main response that feel like a real person thinking, rather than a system outputting

Performance numbers OpenAI published:

• 15.2% higher than gpt-realtime-1.5 on Big Bench Audio (audio intelligence benchmark)
• 13.8% higher on Audio MultiChallenge (instruction following benchmark)

The voice quality of gpt-realtime-2 output is functional and clear. It does not match ElevenLabs' prosodic naturalness — the rhythm and emotional colouring that makes AI speech feel most human. For use cases where the intelligence complexity is the hard problem and voice quality is secondary, this is a significant upgrade. For use cases where the voice itself is part of the product experience, the quality gap to ElevenLabs remains meaningful.

Realtime Translate: Live Speech Translation at Scale

Realtime Translate is a streaming speech translation model. It takes spoken audio in 70+ input languages and translates it into 13 output languages in real time — while the speaker is talking, not after they finish.

The "realtime" distinction matters. Most translation pipelines work in chunks: wait for a sentence to end, send the audio for translation, receive the translated text, synthesise the speech, play it. Realtime Translate processes audio incrementally, keeping pace with the speaker. The result is a conversation that feels more like a bilingual human exchange and less like a queued processing system.

Relevant use cases:

• Multilingual customer support. An English-speaking agent handles a Spanish-speaking caller in real time, with translation flowing in both directions.
• International voice applications. A product built for one language market extends to 13 output languages without rebuilding the voice layer.
• Live interview or meeting translation. Not just for voice agents — any application where people speaking different languages need to communicate in real time.

At 70+ input languages and 13 output languages, coverage is competitive but not the broadest available. ElevenLabs supports 70+ languages on its TTS side; for translation specifically, this model is a strong addition to the API toolbox.

Realtime Whisper: Streaming Transcription

Realtime Whisper is a streaming speech-to-text model. Unlike batch Whisper, which processes complete audio files, Realtime Whisper produces live transcription as the speaker talks — word by word, not utterance by utterance.

This is primarily a developer infrastructure improvement. Existing options like Deepgram Nova-3 already provide streaming STT with strong accuracy. Realtime Whisper's advantage is consolidation: developers who are already in the OpenAI API ecosystem can handle transcription within the same platform rather than routing to a separate STT provider and managing an additional billing relationship.

Accuracy is in line with Deepgram Nova-3 on standard English. Specialised STT providers may maintain advantages on domain-specific vocabulary and accented speech.

Building voice apps? ElevenLabs leads on voice quality — try free today

Pricing

OpenAI's pricing for the new voice models:

Model	Billing	Rate
gpt-realtime-2	Per token	$32/M audio input tokens, $64/M audio output tokens
Realtime Translate	Per minute	See OpenAI pricing page
Realtime Whisper	Per minute	See OpenAI pricing page

For reference: 1 million audio tokens is roughly 16 hours of audio. A typical customer support call is 3–5 minutes — placing single-call gpt-realtime-2 costs at fractions of a cent for audio processing, before LLM inference costs on the reasoning layer.

The overall per-minute cost of a complete gpt-realtime-2 voice agent deployment — including audio input tokens, audio output tokens, and any tool call overhead — will be competitive with equivalent GPT-4o + ElevenLabs TTS stacks in Vapi or Retell. OpenAI's pricing calculator gives the clearest current projection for your specific use pattern.

What This Means for Voice Developers

The three releases collectively represent OpenAI completing their stack play in voice. Previously, a developer building a sophisticated voice agent needed:

• OpenAI for reasoning (GPT-4o via text)
• A separate realtime audio layer (Vapi, Retell, or custom WebSocket)
• A TTS provider (ElevenLabs, Cartesia)
• A STT provider (Deepgram)

With gpt-realtime-2 + Realtime Translate + Realtime Whisper, OpenAI offers:

• Reasoning-capable voice-to-voice in a single model
• Live translation built in
• Streaming STT built in

The consolidation argument is real. Fewer providers, one API, one billing relationship. For teams that prioritise stack simplicity, this is compelling.

How This Changes the Market

The clearest effect is on STT providers. Deepgram, AssemblyAI, and Gladia now compete with a streaming Whisper model embedded inside the largest AI API platform in the market. Developers with no strong reason to use a specialised STT provider will increasingly default to Realtime Whisper for simplicity.

The effect on TTS providers like ElevenLabs is subtler. gpt-realtime-2's voice output is functional — but it was never competing on voice quality, and its May 2026 release did not change that. ElevenLabs' voice naturalness, Expressive Mode, voice cloning, and the full ElevenAgents platform are not replicated by an API model. The competitive pressure is at the developer-convenience layer, not the quality layer.

For voice agent platforms (Vapi, Retell), the consolidation creates a new question: if OpenAI handles reasoning, translation, and STT natively, what is the orchestration layer for? The answer involves telephony, analytics, provider redundancy, and deployment tooling — real value — but the surface area narrows.

What to Do With This Information

If you are a developer using Vapi or Retell with GPT-4o + Deepgram + ElevenLabs: Run a benchmarked comparison of gpt-realtime-2 against your current stack on real calls. The reasoning upgrade in gpt-realtime-2 may improve call quality for complex use cases. Retain ElevenLabs TTS unless the consolidated stack's quality is acceptable for your product.

If you are building a new multilingual voice application: Realtime Translate is worth evaluating before assembling a custom translation pipeline. The 13 output languages may not cover all your markets, but for common language pairs it removes meaningful infrastructure complexity.

If you are a business owner considering a voice agent (not a developer): This announcement does not change the practical landscape for you. ElevenAgents remains the clearest path to deploying a natural-sounding voice agent without engineering resource.

ElevenAgents: deploy a business voice agent in under an hour — no developer needed

Published May 2026 based on OpenAI's release announcements. Model availability and pricing subject to change — check OpenAI's developer documentation for current details.