On September 10, 2026, OpenAI released GPT-Live-1 in its API, giving developers a full-duplex voice model that listens and speaks simultaneously while delegating deeper reasoning to backend models such as GPT-6 Astra. The model replaces the traditional chained STT–LLM–TTS pipeline with a single model that handles interruption, backchanneling, and turn-taking in real time.
GPT-Live-1 first appeared in ChatGPT on July 31, 2026, where it used GPT-5.5 as its reasoning backend. The API version lets developers choose their own backend model and agent harness, pairing the front-end voice layer with models like Luna for high-volume tasks or Astra for complex reasoning. OpenAI says the model cuts interruptions by nearly 80% versus turn-based systems, citing early evaluations with language-learning app Speak.
Confirmed
- Availability: GPT-Live-1 is live in the API as of September 10, 2026 (https://developers.openai.com/api/docs/live).
- Pricing: $0.05 per minute for the front-end voice layer; backend model costs are separate.
- Architecture: Full-duplex, single-model audio I/O with native ASR transcripts, turn detection, alphanumeric understanding, and keyword biasing.
- Delegation: Can hand off reasoning and tool calls to a backend text model (e.g., GPT-6 Astra, third-party models) while keeping the conversation flowing.
- Customization: Tone, pace, and conversational style controllable via system prompt; expanded voice library across accents, dialects, and languages.
- Telephony: Supports full-duplex voice agents for phone calls (reservations, customer support).
- Safety: Audio generated via ChatGPT Voice and the API now includes SynthID watermarking (added July 31, 2026); verification tool and API available at https://openai.com/research/verify/.
- Benchmarks (vendor-reported): OpenAI says GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1; paired with Astra at medium reasoning effort it ranks #1 on Tau3 for end-to-end voice-agent intelligence across airline, retail, and telecom domains.
Unknown
- Independent replication of the 30-point Full Duplex Bench gain and Tau3 #1 ranking.
- Latency SLA, maximum concurrent sessions, and regional availability for the API endpoint.
- Exact model versions behind the “GPT-6 Astra” and “Luna” labels — OpenAI has not published model cards for these backend identifiers.
- Pricing for custom voice access (contact sales) and enterprise deployment via OpenAI Presence.
- Whether the 80% interruption reduction from Speak’s evaluation generalizes beyond language-tutoring workloads.
Our take
GPT-Live-1 moves the conversational layer into a single API call, letting teams swap reasoning backends per task without rewiring audio plumbing. The 30-point Full Duplex Bench gain and Tau3 lead are promising but unreplicated; treat them as directional. At $0.05/minute for the voice layer alone, total cost per agent job will hinge on backend token usage and tool-call volume — list price per token is not total cost per task.
