$ cat wiki/models/gpt-realtime-2.md
GPT-Realtime-2 (OpenAI)
Spec
| Attribute | Value |
|---|---|
| Developer | OpenAI |
| Released | 2026-05-07 (Realtime API GA) |
| Announced | 2026-05-07 |
| Context window | 128,000 tokens (4× increase from GPT-Realtime-1.5's 32K) |
| Pricing | unknown |
| License | unknown |
| Availability | Realtime API (generally available for production) |
| Reasoning effort | Adjustable — minimal / low (default) / medium / high / xhigh |
| Type | Realtime voice model with GPT-5-class reasoning |
Release Date
2026-05-07. Simultaneously: Realtime API exits beta → generally available for production.
Models in Suite
Three models released together:
| Model | Function |
|---|---|
| GPT-Realtime-2 | Voice reasoning (GPT-5-class), tool-calling, interruption handling |
| GPT-Realtime-Translate | Live translation: 70+ input → 13 output languages |
| GPT-Realtime-Whisper | Streaming STT: transcribes live as speaker talks |
Benchmarks
- Big Bench Audio (audio intelligence): GPT-Realtime-2 high = +15.2% vs GPT-Realtime-1.5
- Audio MultiChallenge (instruction following): GPT-Realtime-2 xhigh = +13.8% vs GPT-Realtime-1.5
Key Capabilities
- Simultaneous tool calls: Can call multiple tools at once while speaking ("checking your calendar")
- Interruption handling: Stays responsive while reasoning through complex requests
- Adjustable latency/intelligence tradeoff: xhigh for complex requests, minimal for quick responses
- 128K context: Supports long voice-based workflows (call center, medical documentation, extended sessions)
Use Cases
- Production voice agents (first time GA)
- Customer support voice bots with GPT-5 reasoning
- Medical transcription and documentation
- Live translation services (meetings, conferences)
- Multilingual customer service
Compared To
| Model | Reasoning | Context | Voice | GA |
|---|---|---|---|---|
| GPT-Realtime-2 | GPT-5-class | 128K | Yes | Yes (2026-05-07) |
| GPT-Realtime-1.5 | GPT-4-class | 32K | Yes | No (beta) |
| Grok Voice API | — | — | Yes | Yes |
| Gemini Live | — | — | Yes | Yes (I/O 2026) |
Significance
- OpenAI's first voice model with GPT-5-class reasoning
- Realtime API GA: voice agents now production-ready (previously beta/experimental only)
- Adjustable reasoning effort for voice matches OpenAI's text model API pattern — unified UX
- Moves realtime audio from "call-and-response" to "agentic voice interface"
- GPT-Realtime-Translate: 70+ input languages for live translation is the broadest coverage in market
Open Questions
- How does reasoning latency affect perceived naturalness in conversation?
- Will reasoning effort be auto-selected in future (like o3's auto mode)?
- Real-world accuracy of live translation for low-resource languages?
Sources
- (source)
Related
- OpenAI
- GPT-5.5 Instant — text reasoning counterpart
- Reasoning Models — voice reasoning = reasoning-models entering new modality