Yelp is putting AI voice chat to work on phone reservations. A caller and a software voice can speak at the same time without the model having to stop listening. OpenAI is opening GPT-Live-1 to developers through an API (an interface for connecting software), THE DECODER reports. ChatGPT already uses the model. The release gives developers access to the same simultaneous listening and talking capability for their own AI voice conversations. [1]
AI Voice Chat Takes Both Sides
GPT-Live-1 handles incoming speech while producing its own spoken response, a capability known as full-duplex (simultaneous communication in both directions), rather like a conversation in which neither speaker must surrender the microphone. A caller can interrupt. Thomas Claburn’s September 10 headline in The Register also identifies simultaneous talking and listening as the feature OpenAI is bringing to developers. Matthias Bastian’s report in THE DECODER supplies the model name and details. [1, 2]
Developers can choose the backend model. THE DECODER reports that teams can pair GPT-Live-1 with different models according to their reasoning, speed and cost requirements. Matthias Bastian does not name the available combinations. OpenAI gives developers a choice behind the voice. A team building a reservation service can weigh those requirements separately from its choice of speech interface. Bastian supplies no performance comparison between individual backend configurations. OpenAI Developers says conversation can continue while the backend handles reasoning and tool calls. AIincider and Singularity.Kiwi named GPT-5.5 as the background model for the earlier ChatGPT launch; their July reports do not establish the September API options. [1, 3, 4, 5]
Yelp provides the first concrete job in the account: phone reservations. Yelp chief technology officer Alex Levy reports better call handling. THE DECODER gives no numerical result for the deployment or details about the calls involved. No customer success rate appears. Yelp’s experience brings AI voice chat into a working service; OpenAI’s benchmarks examine how the model behaves when speech overlaps and the conversational turn changes. [1]
OpenAI Reports Faster Turn Taking
OpenAI reports an 80.1 percent score for full-duplex interactivity, compared with 45.4 percent for GPT-Realtime-2.1. THE DECODER attributes the conversational interaction benchmark to OpenAI. Speed gets a separate measurement. Turn-taking latency (the delay when speakers exchange turns) falls from 1.4 seconds to 0.8 seconds, according to the same report. Those figures describe the handover between speakers; they do not time a completed reservation. [1]
The reported delay shrinks by 0.6 seconds. OpenAI also tests tool use, moving the comparison beyond the flow of speech. [1]
OpenAI’s interactivity and timing results measure different parts of a call. A faster handover cannot establish whether an application correctly understands a request and carries it through to completion. The supplied account omits the full test protocol. It offers no independent replication, either. For AI voice chat developers, the next question concerns the software behind the conversation: can GPT-Live-1 use its connected tools accurately while handling the exchange? [1]
Tool Use Faces a Harder Test
OpenAI reports 87 percent tool-calling accuracy for GPT-Live-1, up from 60 percent for the earlier model in the comparison. THE DECODER attributes both figures to OpenAI. Tool calling connects conversation to action. A model uses it to request operations from other software. The report provides no breakdown of individual tools or errors, leaving developers without a task-by-task account of where GPT-Live-1 gained ground. [1]
OpenAI’s banking voice support benchmark sets a harder hurdle: GPT-Live-1 passes 32 percent of cases, compared with 12.4 percent for the previous model, while still falling short in most cases. Fewer than one in three pass. Developers cannot treat the 87 percent tool-calling score as the success rate for an entire support conversation. THE DECODER presents separate benchmarks without establishing that they cover the same tasks. OpenAI supplies both figures. The results have not been independently verified in the supplied reporting, and Yelp’s account of better reservation calls cannot validate a banking test involving a different service and task. [1]
Software that takes actions also features in PerEXP Teamworks’ coverage of Meta Muse’s shopping tools and privacy questions. GPT-Live-1 brings a voice interface to that broader subject of agents carrying out requests. Yelp’s reservations offer a concrete application, but the banking benchmark leaves a gap between conversational fluency and dependable completion. What else does OpenAI give teams to build around that voice? [1]
Twelve Voices, Plus Written Output
THE DECODER reports that GPT-Live-1 offers twelve new voices, with variation in language, dialect and accent. Matthias Bastian does not name the voices or specify their languages. Written output comes built in. Developers receive ASR transcripts (automatic speech recognition text) and response text, giving applications written material from the exchange as well as audio. The report does not specify the timing or formatting of those text streams. [1]
THE DECODER gives no transcript-accuracy score. Its account describes the written outputs as standard features rather than separately demonstrated improvements. [1]
Developers choosing among the twelve voices still have to choose a backend model. Reasoning depth, speed and cost all shape the GPT-Live-1 configuration THE DECODER describes. The report names no preferred pairing. It also supplies no evidence that each combination produces the same benchmark results. An AI voice chat service will need those configuration details to connect OpenAI’s published measurements with the particular setup it runs. [1]
The Per-Minute Price Needs Context
THE DECODER quotes $0.05 per minute for GPT-Live-1. The supplied material does not support that price with an official OpenAI pricing page or detailed billing conditions. Sixty billable minutes would cost $3. That is simple multiplication of the quoted rate, not a verified total for operating an application. The account does not establish whether separate backend processing or other service charges would increase the bill. [1]
Yelp’s reservation calls give OpenAI an example of the speech model doing a specific job. The API release lets other developers build services around the capability already running in ChatGPT. Independent performance evidence remains absent from the package. THE DECODER combines OpenAI’s measurements with Alex Levy’s report of better call handling; neither supplies the detailed customer deployment data needed to compare success across reservation and support applications. [1]
THE DECODER directs developers to OpenAI’s API documentation for the full details. Its short report omits the supported backend list, voice identities and billing conditions. For teams evaluating AI voice chat, those are practical gaps: which configuration matches the calls a service handles, and what does that configuration cost once the conversation starts?
- ONLINE NEWS Bastian, M. (2026, September 10). OpenAI’s GPT-Live-1 API lets developers build apps that talk and listen at the same time. THE DECODER. [Article Link]
- ONLINE NEWS Claburn, T. (2026, September 10). OpenAI arms devs with AI conversation tool that can talk and listen at the same time. The Register. [Article Link]
- WEBSITE OpenAI Developers [@OpenAIDevs]. (2026, September 10). GPT-Live-1 handles listening and speaking in one model [Post]. X. [Article Link]
- ONLINE NEWS AIincider. (2026, July 9). OpenAI’s GPT-Live makes ChatGPT voice talk like a person. [Article Link]
- ONLINE NEWS Singularity.Kiwi. (2026, July 9). OpenAI’s GPT-Live can finally talk and listen at the same time — and it changes everything. [Article Link]
2 comments