Reducing the sequential STT–LLM–TTS chain
OpenAI launched GPT‑Live‑1 in the API on September 10. It jointly processes incoming and outgoing audio so a user can interrupt, pause or change direction while the system is speaking. The voice layer can keep the conversation moving while delegating deeper reasoning and tool calls to GPT‑6 Astra or another backend model.
The release includes prompt control over tone, pace and style, background-noise and silence handling, longer sessions, telephony, ASR transcripts and keyword biasing. The front-end voice layer costs $0.05 per minute; backend model and agent-harness usage is separate. OpenAI reports a 30-percentage-point improvement over GPT‑Realtime‑2.1 on its Full Duplex Bench.
The per-minute price is not the full call cost
The headline price covers the voice front end. Reasoning, tools, telephony, storage and observability can add cost. Speak's interruption result and customer code-reduction claims are early vendor and customer reports, not guarantees across languages, accents, noise profiles and workflows.
Production phone agents still need recording and transcription notice, sensitive-data redaction, speaker verification, explicit confirmation of amounts and addresses, and a reliable path to a human operator.