GPT-Live-1.1 ships in ChatGPT Voice: test before the API flip

OpenAI shipped GPT-Live-1.1 to ChatGPT Voice on October 15, and said API access is "coming soon." The update adds 13 new voices and better performance in French, German, Hindi, Mexican Spanish, Korean and Japanese. If you run a phone or voice agent on GPT-Live-1 in the API, this is your upgrade notice. You have time to test before it reaches you. Use it.

What actually happened

Per OpenAI's developer community announcement and the GPT-Live-1.1 system card:

• Two models. GPT-Live-1.1 replaces GPT-Live-1. A smaller GPT-Live-1.1 Quip replaces GPT-Live-1 mini.
• Who gets which. GPT-Live-1.1 is the default for Plus and higher tiers, including in Codex and Dots. Quip is the default for all other users.
• What changed. 13 new voices, better accents in several non-English languages, and more natural delegation and steerability, per OpenAI.
• Full duplex. Like GPT-Live-1, the models listen and talk at the same time. They decide whether to answer or keep listening when the caller pauses or interrupts.
• The API. Not yet. The announcement says "coming soon to the API," with no date.
• The safety numbers. The system card reports parity with GPT-Live-1 on production voice prompts. On synthetic test prompts it shows statistically significant regressions in some sensitive categories, including mental health and emotional reliance.

OpenAI also wrote that continuous voice model updates "will become the norm going forward." In the community thread, some users report that familiar voices sound different after the update. Those are user reports, not confirmed by OpenAI. They still show what a model update feels like to the person on the other end of the call.

Why it matters for your business: a voice update changes your customer experience

A text model update changes words. A voice model update changes the accent, the pace, the timing of interruptions, and the voice your regular customers know. You notice that the first time a caller hangs up.

Before GPT-Live-1.1 reaches the API, here is what we would do:

• Pin the model version. Call the exact model name. Do not use an alias that moves when OpenAI ships.
• Record a test set now. Twenty real call scripts on GPT-Live-1: a booking, a refund question, a caller who interrupts, a caller with an accent. Run the same set on 1.1 the day it ships, and compare.
• Test your sensitive paths. The system card shows regressions in some sensitive categories. If callers sometimes reach you upset, test those scripts first, and keep your human handoff working.
• Re-check the price. GPT-Live-1 is billed per minute. Confirm the 1.1 price before you switch.

Sources: OpenAI Developer Community: GPT-Live-1.1 is rolling out, OpenAI: GPT-Live-1.1 system card.