GPT-Live 1
Top scores on hard voice agent tasks, but blind listeners prefer Gemini 3.8 Live, which also costs less.
Should you use it?
Pick it if
Developers building voice agents that must finish multi-step tasks, such as customer service.
Best for: Customer service voice agents that must finish multi-step tasks (Voice chat Top 20)
Skip it if
You want the voice people like best for the lowest price; Gemini 3.8 Live ranks higher and costs less.
You give up: Costs more than Gemini 3.8 Live, which blind listeners also prefer
The short version
- What this is
- OpenAI's real-time voice model: it listens and talks at the same time, handles interruptions, and passes hard requests to a backend model. It powers ChatGPT voice and is in the API.
- Why it matters
- It leads the main voice agent benchmark, but people in blind tests prefer a cheaper rival, so it pays to know which matters to you.
- 4th of 26 in blind tests, 1st on the Speech to Speech Index s1 s2; no score yet, as Artificial Analysis's terms stop us publishing one.
- $0.05 a minute in the API plus backend model costs s3.
- API data is not used for training by default s5.
Good at
- First on the Artificial Analysis Speech to Speech Index, 83.2 with the Astra backend s2.
- 74.5% on the τ-Voice customer service test with the Astra backend, against 68.6% for Gemini 3.8 Live's thinking variant s2.
- Handles interruptions: it listens and speaks at the same time and can hand hard requests to a backend model s4 s7.
Our take
OpenAI launched GPT-Live in ChatGPT on 8 Jul 2026, covered by Reuters and SiliconANGLE s6 s7, and made GPT-Live 1 generally available in the API on 10 Sep s8. It is the strongest voice model on Artificial Analysis's combined Speech to Speech Index, mostly because it completes more customer service tasks s2.
In blind arena votes, though, people prefer Gemini 3.8 Live and two Gemini Flash Live variants s1, and Gemini 3.8 Live costs far less per hour s2. Pick GPT-Live 1 when getting the task done matters more than price. Remember that the $0.05 a minute session fee does not include the backend model s3.
Score
A score from 0 to 100 per job, against everything available today.
New to this? What the scores and labels mean
- 0 to 100
- Our score for one job, like writing or coding, compared with everything else out today. 100 is the best available today. For AI models we build it from public data, such as Arena's blind votes and Epoch AI's independent tests; apps and tools are scored on our own checklist. The same tool can score differently for different jobs, and where two likely ranges overlap, we call it a tie.
- Our verdict
- Use it: a top choice for most people. Worth trying: good, and worth a go. Depends: good for some people or one kind of job. Wait: too early to tell, or not out for everyone yet. Skip: something else does the job better.
- First look
- Just released. We say what it is and whether to bother, and give it a score once public results exist.
- No score yet
- None of the public sources we use covers it yet, or the only evidence is from a source whose numbers we can't use, so we describe it in words instead.
- Best buy
- Nothing cheaper scores as well for that job.
- Best free
- The best score for that job that costs nothing, through a free plan or by running on your own computer.
- Pay as you go
- No monthly plan: you pay for each use, usually through a developer account.
- Checked
- The day we last confirmed the score and the price against the sources. Prices change often, so check the date.
- AI model
- The engine underneath, like GPT or Claude. You use it through an app or, if you code, directly.
- Open source
- Free to download and run on your own computer, so nothing leaves your machine.
Voice and audio
No score yet. The leaderboard this rests on does not let us publish numbers from it, so we describe the evidence instead.
2 sources
Public leaderboardIndependent testArtificial Analysis Speech Agent Arena rank 4 of 26, range 3 to 6, GPT-Live-1 (Sol, low), Elo 1053, 652 samples s1. The 26 entries are about 21 distinct models once effort and backend variants are counted once, so the board clears the 20-model small-board rule. Independent check: Artificial Analysis Speech to Speech Index rank 1 of distinct models, GPT-Live-1 (Astra, medium), 83.2 s2, one step away, so no change.
Alternatives
Nothing we have scored beats it for voice on score or on price, and we have not scored a free option for it yet.
- Gemini 3.8 LiveVoice model–FreeRanks 2nd of 26 with blind listeners and costs far less per hour. Side by side
- gpt-realtime-2.1Voice modelNot reviewed yetOpenAI's older realtime voice model, if you are already built on the Realtime API.
- Eleven v4 TurboVoice model–FreeIf you only need text to speech, not a full voice conversation. Side by side
Evidence
Everything behind the call and the scores, folded so the page stays short. Spotted a wrong price, score or link? Report a mistake.
How we scored this
Voice has no score yet, because none of the public sources we use covers it.
Public leaderboardIndependent test
Reviews are drafted with help from AI models, including Anthropic's Claude. Model scores come only from outside leaderboards and independent evaluations, by a written rule.
No one paid for this review or gave us free access, and we earn nothing if you sign up. How we score
Sources 8
- s1Artificial Analysis Speech Agent Arena leaderboard
- s2Artificial Analysis speech to speech leaderboard
- s3OpenAI API pricing
- s4OpenAI API: GPT-Live 1 model page
- s5OpenAI API: data controls
- s6SiliconANGLE: OpenAI launches GPT-Live voice model series
- s7The Star (Reuters): OpenAI launches GPT-Live voice models that listen and speak simultaneously
- s8OpenAI API changelog
Something here out of date or wrong? Report a mistake and we will check it against the source.
History 2 changes
Every change to a rating, the price or our call, with the reason. Scores are against what exists on the day, so they drift down as better things ship. Older entries come from before we moved to scores.
- New: Voice ratedReviewed: voice.
- Audit: re-opened the Speech Agent Arena; rank 4 of 26, range 3 to 6, Elo 1053 and 652 samples confirmed. Evidence now records the distinct-model count for the small-board rule. Rating unchanged.
Use it well 4 guides on this site
Pages on this site that help with the jobs above.