OpenAI's GPT-Live demo videos make it look like you could point that same voice at a phone line tomorrow and get a customer service agent that talks like a real person. You cannot, not yet, and the gap between what the launch demo shows and what a developer can actually build with today is the part most of the coverage skipped.
You cannot use GPT-Live directly to build an AI voice agent for phone calls right now. GPT-Live launched July 8, 2026 as a ChatGPT Voice feature with no public API. What you can actually build phone-based voice agents with today is OpenAI's separate Realtime API, or platforms like Vapi, Retell, Bland, and ElevenLabs that already build on OpenAI's speech models.
That distinction matters whether you are a developer deciding what to build on, an agency owner fielding a client's question, or an executive who watched a demo video and wants to know how fast this can show up in your call center. Here is what is actually true as of this week.
Is GPT-Live Available for Building AI Voice Agents Right Now?
No. GPT-Live is a consumer feature inside ChatGPT Voice, available on iOS, Android, and ChatGPT.com. It is not something a developer can call from their own code today. OpenAI's own launch post says GPT-Live-1 is planned for the API, and developers and enterprises can sign up to be notified when it opens, which is a clear signal that general access was not ready at launch and has no confirmed date.
This is the single most common misunderstanding driving the question. The marketing video shows a consumer product. The question "can I use this for my voice agent" assumes developer access that does not exist yet.
What Can You Actually Build a Phone-Based AI Voice Agent With Today?
Two real paths exist right now, and neither of them is GPT-Live.
OpenAI's Realtime API, running the gpt-realtime model, has been generally available since August 2025 and directly supports phone calling through Session Initiation Protocol (SIP), which is how a voice agent connects to the public phone network, a PBX system, or desk phones. It processes audio through a single speech-to-speech model rather than chaining separate transcription, reasoning, and speech models together, which is the same category of architecture GPT-Live uses, just without the full-duplex, continuous-listening layer GPT-Live adds on top. It also natively supports G.711, the 8kHz codec traditional phone lines use, so it does not need a separate format conversion step to handle a real phone call.
Managed voice AI platforms like Vapi, Retell, Bland, and ElevenLabs give teams a faster path than calling OpenAI's API directly. They handle telephony, call routing, and orchestration, and most of them already support OpenAI's speech models as one of several provider options underneath their platform. An agency building a phone agent today is almost always building on one of these, not on a raw API.
| Option | What it is | Phone calling support | Available today |
|---|---|---|---|
| GPT-Live | Consumer ChatGPT Voice feature | No | Consumer app only |
OpenAI Realtime API (gpt-realtime) | Developer speech-to-speech model | Yes, via SIP | Yes, since August 2025 |
| Vapi / Retell / Bland / ElevenLabs | Managed voice AI platforms | Yes, built-in | Yes |
How Is GPT-Live Different From the OpenAI Realtime API Developers Already Use?
They solve different problems. The Realtime API is built for developers who need a reliable, low-latency speech-to-speech model they can wire into their own application, including phone calls. GPT-Live is built for a specific consumer experience inside ChatGPT, and the marketing videos are showing real behavior, not just a scripted demo. According to OpenAI, GPT-Live listens continuously while speaking, drops in backchannel cues like "mhmm" or "yeah" mid-sentence, waits instead of interrupting when someone pauses to think, and now filters out background noise like traffic or a nearby conversation better than the version before it. It also hands off anything that needs deep reasoning or web search to GPT-5.5 running in the background, then folds the answer back into the conversation without breaking the flow, and can show visual cards for things like weather, stock prices, sports scores, or a map while you keep talking.
None of that is a voice agent building block yet. It is a fixed, packaged experience running on OpenAI's own predefined voices (nine of them, all remastered for this release), with safety training specifically covering emotionally sensitive situations like self-harm risk or emotional reliance on the assistant, and built-in safeguards to prevent the voices from being used to impersonate a real person. That combination of full-duplex listening, reasoning handoff, and voice-specific safety tuning is not yet part of the Realtime API that developers can access. A voice agent built on gpt-realtime today is a strong, production-ready speech-to-speech system. It is not the same continuous, backchannel-aware experience shown in the GPT-Live demo videos.
What Does It Cost to Build a Voice Agent With What Is Available Today?
OpenAI prices gpt-realtime at $32 per 1 million audio input tokens and $64 per 1 million audio output tokens, a 20% reduction from the previous realtime model. Managed platforms price by the minute instead: Retell's base rate is $0.07 per minute with no separate platform fee, and xAI's newly launched Voice Agent Builder undercuts that at $0.05 per minute all in. None of those numbers involve GPT-Live, because GPT-Live is not billed or accessed that way yet.
When Will GPT-Live (or Something Like It) Reach Phone-Based Voice Agents?
There is no committed date. The pattern from OpenAI's last major voice release is worth knowing: the Realtime API launched in beta, then took roughly ten months to reach general availability with production features like SIP calling. If GPT-Live follows a similar arc, developer access is a matter of months, not weeks, and the version that eventually reaches the API may not have every consumer-facing behavior from the ChatGPT demo.
The more useful question for anyone planning around this is not "when does GPT-Live ship" but "what changes about buyer expectations right now." What full-duplex voice AI actually means, and how it compares to the cascaded and turn-based systems most phone agents run on today, covers that shift in more depth.
Frequently Asked Questions
Can Vapi, Retell, or Bland use GPT-Live?
Not yet. These platforms build on models that are available through an API, and GPT-Live is not one of them at launch. Several of them already support OpenAI's Realtime API and gpt-realtime model as one of multiple provider options, which is the closest thing currently available to what GPT-Live demonstrates.
Is GPT-Live the same thing as the OpenAI Realtime API?
No. GPT-Live is a consumer ChatGPT Voice feature with a full-duplex architecture and no public API. The Realtime API is a separate, developer-facing product that has supported phone calling since August 2025. They are both speech-to-speech systems from OpenAI, but they are different products on different release tracks.
Should an agency tell a client they can build them a GPT-Live voice agent?
Not accurately, not today. An agency can build a strong phone-based voice agent using the Realtime API or a managed platform, and can be honest that it uses OpenAI's underlying speech technology. Promising a client "GPT-Live" specifically sets an expectation that cannot be delivered until OpenAI opens API access, which has no confirmed date yet.
Will full-duplex voice AI eventually replace cascaded platforms like Vapi and Retell?
Not entirely, and not soon. Full-duplex is not yet proven at scale over compressed phone audio, and cascaded pipelines remain easier to inspect and control for compliance-heavy call types. The more likely path is that full-duplex becomes available as one option among several, the same way gpt-realtime sits alongside cascaded pipelines today rather than replacing them outright.
Whichever model ends up under the hood, the operating layer around it, routing, reporting, and the ability to switch providers without a rebuild, is what keeps a voice AI deployment stable while the underlying technology keeps changing this fast. See how that operating layer works.