OpenAI moves the reasoning model off the voice agent's critical path
GPT-Live-1 costs $0.05 a minute and passes slow requests to a backend model through a delegation event. One early customer reports deleting 23,000 lines of orchestration, about 80 percent of its voice codebase.
The Engineer · Build desk
What happened
- OpenAI launched GPT-Live-1 in its API on Wednesday, giving outside developers the native full-duplex voice architecture that already runs behind ChatGPT voice mode.
- The model listens and speaks at once, handling interruptions and acknowledgments as they happen while deeper reasoning runs through paired models and tools such as GPT-6 Astra, Codex and ChatGPT Work.
- When a request needs more processing, the voice model hands it off in the background to GPT-6 Astra, to a smaller model like Luna, or to something from another provider entirely.
- Twelve voices ship at launch, spanning accents, dialects and languages, and custom voice access requires contacting sales.
Why it matters
- cost The per-minute voice meter runs for the whole call and every delegation adds a second bill, so what an agent costs now tracks how often it reaches for a reasoning model.
- decision Teams that built their own barge-in and turn-taking glue have to decide whether to keep maintaining it or hand turn control to a hosted model whose timing behaviour they can observe but not change.
- constraint OpenAI's top-line ranking depends on pairing the voice model with GPT-6 Astra at medium reasoning, so a team routing cheap turns to Luna to save money is not running the configuration that was measured.
- exposure Yelp has this in front of live restaurant callers, according to CTO Alex Levy, so a wrong tool call now lands on someone's booking instead of a test transcript.
When GPT-Live-1 needs a bigger model, the voice session generates a `delegation_id`, sends context to whatever backend is handling the heavier work, and gets the result back through an event called `session.commentary.append` [19]. The voice model then folds that result into the conversation instead of reading a block of text aloud [20]. While the backend runs, the voice model holds the turn, filling pauses and acknowledging the speaker, and works the answer in when the backend finishes [6]. OpenAI says processing incoming and outgoing audio together avoids the latency and fragile handoffs that lose timing, context and conversational rhythm [17].
Developers still pick the backend model and the agent harness, and still get ASR transcripts, response text, keyword biasing, alphanumeric recognition and native turn detection [3]. What the deletion leaves is policy. They can still see what the model hears and says, and control when it takes a turn [2]. OpenAI's API docs walk through the full pattern with a working Codex SDK example [28]. The choice of example is a fair hint about the harness they expect you to reach for.
The Decoder reports the full-duplex interactivity score at 80.1 percent against 45.4 percent for GPT-Realtime-2.1 [22], a gap of 34.7 points. OpenAI's own headline figure is 30 percentage points on Full Duplex Bench [10], so the two published numbers are not measuring the same thing [26]. Both come from OpenAI's tests [22].
Tool-calling accuracy is 87 percent, up from 60 percent [24], which leaves 13 calls in 100 wrong [32]. On a banking voice support benchmark the pass rate is 32 percent, up from 12.4 percent [25]; roughly two thirds of that suite still fails [31]. Speak's result came from early tests of its Live Tutor Lessons, where the model was nearly 80 percent less likely to interrupt someone who had simply paused to think [15]. For that to transfer, your callers have to pause mid-sentence the way language learners do.
Tony Stoyanov, co-founder and CTO of EliseAI, said the switch shrank his codebase by 80 percent [13], and the count removed was 23,000 lines [12] running real-time patient conversations [14]. That puts the pre-migration voice stack near 28,750 lines and what remains near 5,750 [27]. The other named early users are Yelp Host, Speak, Intercom's Fin and Cognition's Devin [8]; EliseAI is the only one of them with a published count.
The voice layer bills at $0.05 per minute, about $3 an hour [4], with backend model and agent harness fees charged separately [5]. A ten-minute reservation call is 50 cents of voice before a single reasoning token is billed [21], and telephony is exactly where OpenAI is aiming this: reservations, order updates and customer service [29].
What to watch
- Whether Intercom's Fin or Cognition's Devin publishes its own line count or latency figure, which would tell you if EliseAI's 80 percent is typical or a one-off.
- Whether OpenAI releases the Full Duplex Bench and banking benchmark methodology in enough detail to reproduce the pass rates outside its own harness.
- Whether the promised additional voices and languages arrive with pricing attached, and what custom voice access costs once sales is involved.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption35
- Hype gap+20
- Incentives70
- Confidence60
Perspective Coverage
3 publishers- Builder
- Builder 51%
- Operator
- Operator 32%
- Investor
- Investor 17%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
On Wednesday OpenAI launched GPT-Live-1 in its API, bringing the native, full-duplex voice architecture behind ChatGPT voice mode to outside developers for the first time.
- [2]
Developers can still see what the model hears and says and control when it takes a turn.
- [3]
Developers can control tone, pace and style through the system prompt, select their own backend model and agent harness, and use ASR transcripts, response text, keyword biasing, alphanumeric recognition and native turn detection.
- [4]
GPT-Live-1 costs $0.05 per minute, or about $3 an hour.
- [5]
Backend model and agent harness fees are charged separately from the $0.05 per minute voice price.
- [6]
GPT-Live-1 can keep the conversation going while a backend model works, filling pauses and acknowledging the speaker, then work the answer in once the backend is finished.
- [7]
When a request needs more time or more processing, GPT-Live-1 can hand it off to another model in the background: GPT-6 Astra, a smaller model like Luna, or something from another provider entirely.
- [8]
Early users include Yelp Host, Speak, Intercom's Fin, and Cognition's Devin.
- [9]
Yelp CTO Alex Levy said the company is seeing more calls successfully handled by AI, and callers are speaking in fuller, more natural sentences.
- [10]
OpenAI says GPT-Live-1 performs 30 percentage points better than GPT-Realtime-2.1 on Full Duplex Bench.
- [11]
Paired with GPT-6 Astra at medium reasoning, GPT-Live-1 takes the top spot on the tau-cubed benchmark.
- [12]
One early customer deleted 23,000 lines of code after switching to GPT-Live-1.
- [13]
Tony Stoyanov, co-founder and CTO of EliseAI, a healthcare company testing the API, said the move shrank his codebase by 80 percent.
- [14]
One customer said the move from a cascaded system cut its voice codebase by 80 percent and removed 23,000 lines used for real-time patient conversations.
- [15]
In early tests of Speak's Live Tutor Lessons, GPT-Live-1 was nearly 80 percent less likely to interrupt someone who had simply paused to think.
- [16]
GPT-Live-1 launches with 12 voices spanning accents, dialects and languages, and custom voice access requires contacting sales.
- [17]
OpenAI says processing incoming and outgoing audio together avoids latency and fragile handoffs that can lose timing, context, and conversational rhythm.
- [18]
The model listens and speaks at the same time, handles interruptions and acknowledgments as they happen, and can keep a conversation moving while deeper reasoning or actions run through paired models and tools such as GPT-6 Astra, Codex, and ChatGPT Work.
- [19]
OpenAI exposes delegation through an event-driven interface: the voice session generates a delegation_id, sends context to whatever backend system is handling the heavier work, and gets the result back through an event called session.commentary.append.
- [20]
The voice model folds the delegated result into the ongoing conversation rather than reading a block of text aloud.
- [21]
At $0.05 per minute, a ten-minute call costs 50 cents for the voice layer alone.
- [22]
On OpenAI's benchmarks, in full-duplex interactivity tests GPT-Live-1 scores 80.1 percent compared with 45.4 percent for GPT-Realtime-2.1.
- [23]
Turn-taking latency drops to 0.8 seconds from 1.4 seconds.
- [24]
Tool-calling accuracy jumps to 87 percent from 60 percent.
- [25]
In a banking voice support benchmark, GPT-Live-1 hits a 32 percent pass rate, up from 12.4 percent for the previous model.
- [26]
The interactivity scores reported by The Decoder differ by 34.7 points, which is larger than the 30 percentage points OpenAI reports for Full Duplex Bench, so the two published figures are not the same measurement.
- [27]
If 23,000 lines were 80 percent of the voice codebase, the codebase was about 28,750 lines before the migration and about 5,750 lines after.
- [28]
OpenAI's API docs walk through the full delegation pattern, including a working example with the Codex SDK.
- [29]
Telephony support targets reservations, order updates, and customer service.
- [30]
OpenAI plans to add more voices and languages, and points enterprises to Presence for real-time workflows that can use company systems, take approved actions, and escalate to people.
- [31]
A 32 percent pass rate on the banking voice support benchmark leaves 68 percent of that suite failing.
- [32]
Tool-calling accuracy of 87 percent means 13 calls in 100 are wrong.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
3 independent publishers whose own reporting we read for this story.
- testingcatalog.comOpenAI launches GPT-Live-1 for full-duplex voice agents
1 article · September 10, 2026
- the-decoder.comOpenAI's GPT-Live-1 API lets developers build apps that talk and listen at the same time
1 article · September 10, 2026
- thenewstack.ioOpenAI split a voice model’s brain. Then one team deleted 23,000 lines of code.
1 article · September 10, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Voice agents and live captioningFollow
- Full-Duplex Speech ModelsFollow
- AI API pricingFollow