Sign in Subscribe
news

2026-05-07 GPT-Realtime-2 - OpenAI's voice model gets GPT-5-class reasoning

Sebastien Dubois

Sebastien Dubois

— 5 min read

Canonical version: 2026-05-07 GPT-Realtime-2 - OpenAI's voice model gets GPT-5-class reasoning.

On 7 May 2026, OpenAI released three new audio models in its Realtime API: GPT-Realtime-2, GPT-Realtime-Translate and GPT-Realtime-Whisper. OpenAI pitched them as voice agents that can reason, call tools, translate and transcribe while the conversation keeps going.

Until then, OpenAI's voice models lagged well behind its text models. Simon Willison pointed that out a month earlier: ChatGPT's voice mode still ran on a GPT-4o era model with a 2024 knowledge cutoff. GPT-Realtime-2 is OpenAI's first voice model with what it calls "GPT-5-class reasoning".

What shipped

Model What it does Endpoint Price
gpt-realtime-2 Speech-to-speech voice agent with reasoning and tools v1/realtime Audio: $32 in, $0.40 cached, $64 out per 1M tokens
gpt-realtime-translate Live speech translation, 70+ input languages to 13 output languages v1/realtime/translations $0.034 per minute
gpt-realtime-whisper Streaming speech-to-text v1/realtime/transcription_sessions $0.017 per minute

No new text-to-speech model came with this release. Speech output comes from GPT-Realtime-2 itself (and from Translate for translated audio).

Changes in GPT-Realtime-2

The main changes compared with GPT-Realtime-1.5:

  • Adjustable reasoning effort: minimal, low (the default), medium, high and xhigh. Low keeps latency down; higher levels take more time and more output tokens
  • Preambles: short phrases like "let me check that" while the model works, so the user doesn't sit in silence
  • Parallel tool calls, made audible ("checking your calendar")
  • Recovery: it says "I'm having trouble with that right now" instead of failing silently
  • Context window from 32K to 128K tokens, and max output from 4,096 to 32,000 tokens (per the model docs)
  • Better handling of specialized terms, proper nouns and healthcare vocabulary
  • More control over tone: calm, empathetic or upbeat depending on the moment

Inputs are text, audio and images; outputs are text and audio. The knowledge cutoff is still 30 September 2024, the same as GPT-Realtime-1.5.

Benchmarks

OpenAI published two comparisons against GPT-Realtime-1.5 (the values come from OpenAI's chart, as reproduced by The Decoder):

  • Big Bench Audio (audio reasoning), at high: 96.6% vs 81.4%
  • Audio MultiChallenge (multi-turn instruction following), at xhigh: 48.5% vs 34.7%

The post describes these as "15.2% higher" and "13.8% higher". Those are percentage points. In relative terms, that's about 19% and 40% better.

Customer claims from the launch post: Zillow says call success on its hardest adversarial benchmark went from 69% to 95% after prompt optimization. BolnaAI says GPT-Realtime-Translate had 12.5% lower word error rates than any other model it tested in Hindi, Tamil and Telugu. Deutsche Telekom, Priceline and Vimeo are also named as early users.

Pricing

Audio pricing for GPT-Realtime-2 is the same as GPT-Realtime-1.5 ($32 / $64 per million audio tokens). Text output went up from $16 to $24 per million tokens. Translate and Whisper are billed per minute of audio, which makes them a lot easier to estimate than token-based voice. (The first question on Hacker News asked how much audio 1M tokens is; it got no answer.)

Availability

All three models were available on launch day in the Realtime API and the Playground. The Realtime API supports EU data residency. OpenAI runs classifiers over sessions and can halt conversations that break its policies, and developers must tell users when they're talking to an AI unless it's obvious.

ChatGPT didn't get anything that day. OpenAI's own reply under its launch tweet: "We know you're eager for voice updates in ChatGPT. Stay tuned, we're cooking."

What the community said

The Hacker News thread stayed small (36 points, 5 comments). Commenters raised two points:

  • GPT-Realtime-Whisper is not open-weight, unlike the original Whisper
  • The talking model can delegate the reasoning. One commenter was already wiring subagents, TTS and STT together in Claude Code, and argued that the voice model only needs to stay present and delegate. That's pretty much the design OpenAI shipped two months later with GPT-Live

What happened since

Timeline from OpenAI's API changelog:

  • 6 July 2026: GPT-Realtime-2.1 (better alphanumeric recognition, silence and noise handling, interruptions; same price) and GPT-Realtime-2.1 mini ($10 / $20 per million audio tokens). OpenAI also says p95 latency dropped by at least 25% across Realtime voice models
  • 8 July 2026: GPT-Live replaces the old model in ChatGPT voice mode, delegating hard tasks to GPT-5.5 in the background
  • 28 July 2026: gpt-live-transcribe (streaming, $0.017 per minute) and gpt-transcribe (files, $0.0045 per minute). The live transcription guide now starts with gpt-live-transcribe, not GPT-Realtime-Whisper
  • 26 August 2026: whisper-1 and the gpt-4o-*-transcribe models are deprecated, shutting down on 26 February 2027
  • 10 September 2026: GPT-Live 1 is generally available in the API at $0.05 per minute. OpenAI's audio guide now recommends it for any new conversational voice app

The deprecations page also lists tts-1, tts-1-hd and the gpt-4o-mini-tts snapshots for removal on 6 January 2027, with gpt-realtime-2.1-mini as the replacement. For text-to-speech, OpenAI now points to a realtime model.

My take

GPT-Realtime-2 closed the gap between the voice model and the text models, at least on paper. Reasoning effort, preambles and audible tool calls are exactly what a voice agent needs. The hard part of a voice product is keeping the conversation alive while something slow (a search, a booking, an API call) happens in the background.

But I wouldn't start a new project on it today:

  • For a new voice agent, look at GPT-Live first. That's what OpenAI's own docs recommend now
  • If you're already on the Realtime API, use gpt-realtime-2.1 (same price as 2) and re-test your function calls. One developer on OpenAI's forum reported that 2.1 mini stopped calling tools in a SIP flow that worked with the older mini
  • For live captions, compare gpt-live-transcribe with GPT-Realtime-Whisper; same price per minute, and the docs now point to the former
  • For live translation, GPT-Realtime-Translate is still the dedicated option. Plan one session per target language
  • None of these are local or open models. For speech-to-text, look at Whisper, Parakeet V3 or Voxtral Transcribe 2; for speech output, Supertonic runs on-device, and my Text-to-Speech (TTS) note lists more options

That's it for today! ✨

References

About Sébastien

Ready to get to the next level?

Found this valuable? Share it with someone who needs it.

Join 6,000+ readers. Get practical systems for knowledge & AI. Free.

Subscribe ✨

Free: Knowledge System Checklist

A clear roadmap to building your own knowledge system. Subscribe and get it straight to your inbox.

6,000+ readers. No spam. Unsubscribe anytime.

Subscribe