OpenAI ships three new realtime voice models — GA, GPT-5-class reasoning, 70-language translation

What OpenAI shipped on May 7

Three models, plus the Realtime API exiting beta. The headline shift is from “voice models that talk to you” to voice models that reason while talking to you — and on the infrastructure layer that production apps can actually deploy on.

GPT-Realtime-2

The flagship: a voice model carrying GPT-5-class reasoning. It can carry a conversation forward naturally, handle multi-step requests inside a single voice exchange, ask clarifying questions when context is ambiguous, and call tools (the way text models call tools) without breaking the conversational flow. The practical upgrade over GPT-Realtime-1: you can now ask a voice agent “find the customer’s last three invoices, identify which ones are overdue, draft a reminder email, and read the summary back” — and the model does that as one continuous conversation rather than breaking into discrete API calls.

Pricing is on tokens, not minutes: $32 per million audio-input tokens; $64 per million audio-output tokens. For typical voice-agent workloads (5-10 minute conversations), this lands roughly in the $0.30–0.80 per conversation range — competitive with what production voice-bot teams pay today.

GPT-Realtime-Translate

Live spoken-language translation across 70 input languages → 13 output languages, keeping pace with the speaker (per OpenAI’s own demo, sub-200ms typical latency). Use cases:

• Multi-language customer support without translator infrastructure
• Cross-border B2B sales calls
• Live conference translation for events, presentations, briefings
• Accessibility — real-time captions in a user’s preferred language for any spoken content

Pricing: $0.034 per minute. At that rate, a 30-minute translated meeting costs about $1 — substantially cheaper than hiring an interpreter or running a pre-recorded transcription + translation pipeline.

GPT-Realtime-Whisper

Streaming speech-to-text. The model that powers ChatGPT voice’s “I’m listening” experience is now available as a standalone API: transcription happens live as the speaker talks, not after they finish. This isn’t a Whisper successor in the sense of better accuracy — Whisper-3 already does excellent accuracy. The shift is latency and streaming: transcription tokens arrive as they’re generated rather than after the audio ends.

Pricing: $0.017 per minute. A typical 60-minute meeting transcription costs about $1.02.

The Realtime API exits beta

This is the structural change. The Realtime API has been usable since 2024 but in beta status — production teams have been hesitant to build mission-critical apps on a beta endpoint. General availability removes that gate. Voice agents shipping in production now have a supported, SLA-backed API to build on.

What this means for the voice-AI market

Three competitive shifts to watch:

vs. ElevenLabs

ElevenLabs has owned the voice-AI category since 2024, with two distinct moats:

• Voice cloning quality — IVC and PVC voice clones that are genuinely indistinguishable on first listen
• Creator workflow — Studio, library of voices, multi-language dubbing for video

OpenAI’s May 7 release doesn’t directly attack either moat. GPT-Realtime models don’t ship voice cloning. ElevenLabs’ Dubbing Studio remains the better tool for video creators. But on the conversational voice-agent axis — call-center automation, real-time interactive voice products, multilingual support flows — OpenAI now has the technically strongest offering, integrated with GPT-5-class reasoning that ElevenLabs Conversational doesn’t match.

The market read: ElevenLabs remains the right choice for creators producing voice content (audiobooks, podcasts, narration, dubbed video). OpenAI’s Realtime models are the right choice for developers building voice-first products at scale. Both can coexist; they’re not fully substitutable.

vs. the call-center category (Sierra, Decagon, etc.)

This is where the competitive picture is more interesting. Sierra and Decagon — the leading AI customer-service agent platforms — have built their voice products on top of voice models from OpenAI, Anthropic, ElevenLabs, and others. GPT-Realtime-2’s GPT-5-class reasoning probably gets adopted by Sierra and Decagon as a model option rather than competing with them directly. Sierra and Decagon’s product moats are the workflow definition layer (Agent Operating Procedures, etc.), not the underlying voice model.

What changes: customers evaluating Sierra/Decagon will probably ask “can your platform run on GPT-Realtime-2?” by Q3 2026. Sierra and Decagon will say yes (because the API is there); the competitive shape stays roughly the same.

vs. Google Gemini Live, Apple Intelligence voice

Google’s Gemini Live has had a competitive voice-conversation product for over a year. Apple has been pushing voice updates in Apple Intelligence on the device. The differential here is the API-availability vs consumer-product framing: Google and Apple’s voice products work great inside Google/Apple ecosystems but neither offers a developer API as accessible as OpenAI’s Realtime. For developers building cross-platform voice products, OpenAI remains the easier integration. For consumers, Google + Apple are competing on a different axis entirely.

What it doesn’t change

A few things to keep in proportion:

Quality on extreme emotional dramatic narration is still ElevenLabs territory. GPT-Realtime-2 handles conversational warmth well; it doesn’t yet match human voice actors on dramatic intensity for high-stakes creative work. For audiobooks and dramatic voice-over, ElevenLabs PVC remains the better tool.

Privacy concerns around continuous audio capture remain real. Voice interfaces that listen-continuously raise real surveillance and data-handling questions. OpenAI’s enterprise tier offers data-handling controls; consumer-tier deployments still need careful UX around consent and recording.

Audio quality on lower-bandwidth connections varies. Real-world voice apps deployed across cellular networks see meaningfully more latency and quality variance than OpenAI’s demo conditions. Production tuning is still real work.

What this means for Pick Right readers

If you’re using ElevenLabs for voice work: No immediate change. ElevenLabs’ Creator tier remains the right tool for audiobook, podcast, and video-voice production. The OpenAI release affects API-first voice product builders, not creator-tier subscribers.

If you’re building a voice product with OpenAI: The Realtime API GA matters — production SLA, multiple model options for transcription / translation / reasoning, and pricing that pencils for real workloads. Consider migrating from Whisper-3 batch transcription to streaming-Whisper for any user-facing latency-sensitive use case.

If you’re evaluating ChatGPT voice mode: No direct impact. Consumer ChatGPT voice still runs on the same underlying models; this release is the API surface for developers, not a consumer feature change.

If you’re tracking the broader AI category: This release confirms that voice is becoming infrastructure, not a premium feature. OpenAI shipping a $0.017/minute streaming-Whisper means transcription is essentially a commodity. The “premium voice model” framing of 2024 has shifted; differentiation moves further up the stack into reasoning quality, integration depth, and workflow surfaces.

For broader category context, see the ElevenLabs review, the ChatGPT review, and best AI audio tools.

The pragmatic read

OpenAI ships voice models with GPT-5-class reasoning at infrastructure pricing, exits beta on the developer API, and adds live translation across 70 languages. The market interpretation: voice-AI is no longer a “specialized product category” — it’s an infrastructure capability that any product can adopt for cents per minute. ElevenLabs still leads on creator workflow and voice cloning. The competitive ground is now whether the value sits in the voice model (where OpenAI is competing aggressively) or the voice workflow (where ElevenLabs and the call-center platforms still win).

The right take for most readers: this is structurally important for the voice-AI market but doesn’t change your immediate tool choices. If you’re an ElevenLabs customer, stay there. If you’re a ChatGPT user, the consumer experience didn’t change. If you’re a developer building voice-first products, the May 7 release just made your stack 30% cheaper and 50% more capable.

Questions or corrections? Email Pick Right. Want the full list? See all news.