Published 11 September 2026
OpenAI put GPT-Live-1 into its API on 10 September 2026 at $0.05 per minute of conversation. The model listens and speaks at the same time, which OpenAI says lifts its Full Duplex Bench result by 30 percentage points over GPT-Realtime-2.1. That price covers the voice layer only, so whatever model does the thinking behind it is billed on top.
What is GPT-Live-1 and what does full-duplex mean?
Most voice agents today chain three things together: speech recognition, a text model, and speech synthesis. Every handoff costs time, and the timing of a conversation is easy to lose. GPT-Live-1 replaces that chain with one model that handles incoming and outgoing audio together. Full-duplex is the word for it, and it means the model can keep listening while it is still talking.
The reasoning does not happen inside GPT-Live-1. You pick a text model to sit behind it and do the thinking and the tool calls, and OpenAI names GPT-6 Astra, Luna and Terra as options, alongside third-party models. The idea is to match the depth of thought to the task: something cheap for order updates, something stronger for a complicated complaint. GPT-Live-1 keeps talking while that work runs in the background.
It also does the plumbing you would otherwise build yourself. You get speech transcripts and response text natively, keyword biasing, better recognition of things like order numbers and postcodes, and turn detection if your app still wants explicit turns. Telephony is supported, so the agent can answer a phone.
GPT-Live-1 price and specifications against GPT-Realtime-2.1
You read this release most clearly next to the model it replaces. The pricing changes shape entirely, and one capability goes backwards.
| What is measured | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| How you pay | $0.05 per minute of session | $32 per 1M audio input tokens, $64 per 1M output |
| Listens and speaks at once | Yes | No, turn based |
| Full Duplex Bench (pauses, turn taking, interruptions) | 30 points higher | Baseline |
| Image input | Not supported | Supported, $5 per 1M image tokens |
| Voices at launch | 12, custom voices via sales | Count not published |
| Concurrent sessions, Tier 1 to Tier 5 | 25 to 500 | Capped on requests per minute instead |
The Full Duplex Bench figure is OpenAI’s own evaluation, and it publishes the gap rather than the two scores, so you cannot see where either model starts from. The prices come from OpenAI’s model and pricing pages. The two price rows are not directly comparable, and that is the point: you now pay for time on the line instead of for audio tokens, and the token bill moves to whichever model you put behind it.
What OpenAI is not saying about GPT-Live-1
The $0.05 per minute buys the front end and nothing else. OpenAI gives no worked example of what a finished voice agent costs per call, and that is the number a buyer actually needs. The arithmetic on the voice layer alone is simple enough: a thousand hours of calls comes to $3,000 before the model behind it has read a single token.
The Tau3 result is a ranking for a pair, not for the voice model. GPT-Live-1 comes first on that benchmark when it runs with GPT-6 Astra at medium reasoning effort, which is also the most expensive model OpenAI currently sells. Pair it with something cheaper and you are not buying the benchmark you read about.
One capability goes backwards. GPT-Realtime-2.1 accepts image input. GPT-Live-1 does not, and it does not take video either. If your voice agent needs to look at a photo a customer sends during a call, this model cannot do that job on its own.
There is also no latency figure anywhere. OpenAI’s argument is that removing the handoffs cuts delay, but the model page lists both speed and performance as not specified, and no millisecond number appears in the announcement. The customer results it quotes, including the 80% drop in interruptions at Speak and the 23,000 lines of code removed at another customer, are figures those companies supplied.
What you get on day one
- Price: $0.05 per minute of session, billed per second, with the backend model and tools billed separately
- Inputs and outputs: audio and text only, no image and no video
- Voices: 12 at launch across accents and languages, custom voices only through OpenAI sales
- Endpoint: v1/live/sessions, with streaming and function calling supported
- Limits: 25 concurrent sessions on Tier 1 rising to 500 on Tier 5, and no free tier
- Knowledge cutoff: 31 July 2025
What this means
This is worth testing now if you run phone-based support or sales at volume and your current setup is a chain of three services. Think of a five-person team keeping a reservations or claims line running. The value here is not a benchmark, it is deleting the code that decides what happens when a caller interrupts. One customer says that removal took 23,000 lines out of their codebase. If that is anywhere near your situation, the migration pays for itself in maintenance before it pays for itself in call quality.
It is worth watching rather than adopting if your voice product needs to see anything. Dropping image input is a real step back from GPT-Realtime-2.1, and OpenAI has not said when it returns. This is also not a cheaper product than what you have. It is a differently shaped bill, and whether it comes out lower depends entirely on the backend model you choose. That is the one decision OpenAI has left to you, and the one it has given you no numbers for. The first thing we would measure is cost per completed call with a small model behind it, then the same call with Astra, because the gap between those two is the real price of this release.
For more information, visit the official announcement of GPT-Live-1 on the OpenAI blog.
