GPT-Live vs. Vapi, Retell, and ElevenLabs is the comparison agencies are searching for this week, and it is not an even one, because only three of those four are things a team can actually buy. GPT-Live has no public API. Vapi, Retell, and ElevenLabs do, and one of them already ships the full-duplex-style conversation OpenAI just made famous.
GPT-Live vs. Vapi, Retell, and ElevenLabs comes down to this: GPT-Live cannot be built into a phone agent today because OpenAI has not opened its API. Of the three platforms a team can actually build on, ElevenLabs already markets full-duplex-style turn-taking with sub-second latency, Retell wins on cost and latency for straightforward calls, and Vapi wins on flexibility to swap components underneath.
That makes this less a four-way comparison and more a two-part decision: what full-duplex naturalness is actually available right now, and which of the three real platforms fits a given deployment.
| GPT-Live | Vapi | Retell | ElevenLabs | |
|---|---|---|---|---|
| What it is | Consumer ChatGPT Voice feature | Orchestration layer over chosen models | Cascaded pipeline (STT, LLM, TTS) | Conversational AI 2.0 platform |
| Full-duplex or turn-taking claim | Yes, full-duplex | Depends on components chosen | Turn-detection based, not full-duplex | Yes, markets full-duplex conversation |
| Public API available | No | Yes | Yes | Yes |
| Telephony / phone calling support | No | Yes | Yes | Yes |
| Buildable into a phone agent today | No | Yes | Yes | Yes |
Can You Actually Buy GPT-Live for a Voice AI Deployment?
No. GPT-Live launched July 8, 2026 as a ChatGPT Voice feature. OpenAI says an API version is planned and developers can sign up to be notified, but there is no public access, no telephony integration, and no committed date. Anyone comparing it against Vapi, Retell, or ElevenLabs as a buying option today is comparing a live product against a waitlist.
Does Any Current Platform Already Offer Full-Duplex-Style Conversation?
Yes, and this is the detail most coverage of the GPT-Live launch missed. ElevenLabs shipped Conversational AI 2.0 with what it directly calls full-duplex conversation, sub-second latency on its Flash model, and a turn-taking system that reads cues like "um" and "ah" in real time to decide when to speak or wait. That is functionally the same category of behavior GPT-Live demonstrated, shipped as a buildable product months before OpenAI's consumer launch.
The limitation is control. ElevenLabs' turn-taking model does not expose tuning parameters, so a team gets the naturalness but cannot adjust how aggressively the agent interrupts or waits. Retell and Vapi, by contrast, run more traditional cascaded pipelines with explicit turn-detection settings a team can configure directly, at the cost of the same continuous, backchannel-aware feel.
How Do Vapi, Retell, and ElevenLabs Actually Compare on Price and Control?
| Vapi | Retell | ElevenLabs | |
|---|---|---|---|
| Architecture | Orchestration layer; bring your own STT/LLM/TTS or realtime model | Cascaded pipeline (STT, LLM, TTS) | Conversational AI 2.0, full-duplex-style turn-taking |
| Advertised rate | $0.05/min platform fee | $0.07/min, no platform fee | $0.08/min overage (varies by plan) |
| Realistic all-in cost | $0.15 to $0.40/min once STT, LLM, TTS, and telephony are added | roughly $0.13 to $0.31/min with dependencies included | $0.10 to $0.50/min depending on plan tier and overage |
| Measured latency | Depends on components chosen | About 620ms | Sub-second on the Flash model |
| HIPAA / compliance | Available, plus roughly $1,000/month | Included with a self-service BAA on Enterprise plans | Not a headline feature; verify directly before a regulated deployment |
| Best fit | Teams that want to swap components, including a realtime speech-to-speech model later | Straightforward phone deployments where cost and latency matter most | Deployments where conversational naturalness is the priority and tuning control matters less |
What Changes About the Buying Decision Now That Full-Duplex Is Mainstream?
Nothing about the underlying decision changes. What changes is that "full-duplex" is now a term buyers will use in a sales conversation, and a platform that cannot speak to it plainly looks behind. The actual decision still comes down to three questions: how much does the deployment need to cost per minute, does the call type require inspectable, controllable turn detection for compliance reasons, and does the team want the flexibility to swap in a different underlying model later without rebuilding.
GPT-Live raises the bar for what "natural" sounds like, and ElevenLabs is the closest thing to that bar available to build with today. Retell remains the pragmatic choice when a straightforward, cost-predictable phone deployment matters more than backchanneling nuance. Vapi remains the choice for teams that expect to swap providers, including plugging in OpenAI's own Realtime API once that access is easier to manage, without rebuilding the integration layer around it.
Whichever provider gets picked, the pricing numbers above are advertised rates, not what most teams actually pay once STT, LLM, TTS, and telephony are billed separately on top. Verifying current, all-in pricing directly with each vendor before signing is worth the extra hour, because all three change pricing structure often enough that a number from even a few months ago can be stale.
The provider decision also is not the whole decision. Whichever of these three a team picks, the operating layer around it, routing calls to the right provider per client, keeping reporting consistent, and being able to switch providers later without a rebuild, is a separate problem from which conversation-layer model sounds most natural. That is the layer Voxfra operates at, above the provider choice, not as an alternative to it.
Frequently Asked Questions
Is ElevenLabs' full-duplex the same thing as GPT-Live's full-duplex?
They are the same general category, both process input and generate output continuously rather than waiting for silence, but they are not the same product. ElevenLabs' version is available to build with today and does not expose tuning controls. GPT-Live is not yet available outside ChatGPT Voice, and its behavior, including handoff to GPT-5.5 for deeper reasoning, is not part of any third-party platform.
Can Vapi be used to build a GPT-Live-style voice agent?
Not GPT-Live specifically, since it has no API. Vapi's orchestration model does let a team plug in other speech-to-speech options, including OpenAI's separate Realtime API, which is the closest buildable equivalent to GPT-Live's underlying model family.
Which platform is cheapest for a straightforward phone agent?
Retell has the most predictable pricing for standard call types, with a $0.07/min base and no platform fee, landing around $0.13 to $0.31/min once dependencies are included. Vapi and ElevenLabs can both cost less or more than that depending on component choices and plan tier, so the "cheapest" answer depends on the specific call volume and complexity.
Should an agency wait for GPT-Live API access before choosing a platform?
No. There is no confirmed date for GPT-Live API access, and Vapi, Retell, and ElevenLabs are all shipping and improving now. A platform decision made today can change later; a platform decision delayed indefinitely has a real cost in lost deployment time.
Whichever of these three a team picks, the operating layer around the provider choice, routing, reporting, and the ability to switch providers without a rebuild, stays a separate decision from which conversation model sounds most natural. See how that operating layer works.