OpenAI shipped GPT-Live on July 8, and the demos are genuinely uncanny. The model listens and talks at the same time. It says “mhmm” while you’re mid-sentence. You can cut it off and it just… adjusts, the way a person does. Advanced Voice Mode is gone, replaced everywhere.
Then a few thousand developers went looking for the API and found a waitlist form.
That gap is the whole story. If you’re a ChatGPT user, GPT-Live is a free upgrade that landed in your app without you asking. If you’re building a voice agent this quarter, GPT-Live is currently worth exactly zero to you, and the more useful news dropped two days earlier with almost no coverage. Let’s separate those two things, because most of the posts about this launch don’t.
Full-duplex isn’t a buzzword, it’s a different architecture
Every voice assistant you’ve used until now works in turns. You talk, it detects silence, it transcribes, it thinks, it speaks. Even when the pipeline got fast enough to feel conversational, the turn-taking was structural — the model physically could not be listening while it was speaking.
GPT-Live-1 processes incoming audio and generates outgoing audio concurrently. That’s it. That’s the change. But it cascades into things that were previously impossible to fake:
Backchanneling. The little “mhmm,” “right,” “yeah” noises humans make to signal they’re still listening. A turn-based model can’t produce these, because producing them would mean taking a turn.
Real interruption. Not “detect that the user started talking, then stop the TTS stream.” The model hears you cutting in while it’s speaking and can respond to what you said rather than just halting.
Deliberate silence. GPT-Live can decide to say nothing while you think out loud. Sounds trivial. It isn’t — every previous voice mode treated your pause as a cue to start talking, which is why using them felt like being interrupted by an eager intern.
There’s a second layer that I think is the more interesting engineering decision. GPT-Live doesn’t try to be smart. It’s a conversational front end, and when a request needs real reasoning, search, or agentic work, it delegates to GPT-5.5 running in the background while it keeps the conversation warm. Human evals put GPT-Live-1 well ahead of Advanced Voice Mode on turn-taking, interruption handling, and naturalness, and it posts substantial gains on GPQA and BrowseComp too — but those reasoning gains come from the handoff, not from the voice model itself getting a bigger brain.
That’s a smart split. Voice needs sub-200ms responsiveness; reasoning needs seconds. Trying to get both from one model is how you end up with something that’s either dumb or slow. OpenAI decided to stop trying.
The parts nobody demos
Early users are complaining that GPT-Live is too much. The backchanneling is over-enthusiastic — a stream of “mhmm”s and “yeah”s that reads as attentive for about ninety seconds and as grating forever after. It’s a tuning problem, and OpenAI will probably dial it back, but it’s a good reminder that “more human” and “more pleasant to use” aren’t the same axis.
The multilingual story is also weaker than the launch framing suggests. In OpenAI’s own Hindi translation demo, the model spoke with a heavy American accent and a stilted, bookish phrasing. OpenAI says it’s optimized for “most spoken languages” without specifying which, which is the kind of phrasing you use when the answer is “English, and then it depends.” Video input and full multilingual parity aren’t in this release at all.
Free users get GPT-Live-1 mini. Paid users get the full model. Both replaced Advanced Voice Mode globally, so there’s no opting out of the new personality.
The API reality check
There is no GPT-Live API. There’s a form at openai.com/form/gpt-live-1-in-the-api where you can ask to be told when there is one.
OpenAI hasn’t committed to a date. Some outlets are reporting “weeks, not months” — treat that as a rumor, not a roadmap, and price your plans accordingly. Past OpenAI launches have gone consumer-first and then API, usually via a limited developer preview in batches before general availability. If your Q3 launch depends on GPT-Live being in the API, on a date you can hold a vendor to, you don’t have a plan. You have a hope.
So what do you actually build on? Here’s where the July 8 news cycle buried the lede.
gpt-realtime-2.1 is the thing you missed
On July 6 — two days before the GPT-Live consumer launch, and swamped by it — OpenAI shipped gpt-realtime-2.1 and gpt-realtime-2.1-mini to the Realtime API. This is the tier that’s actually available to you, and it got a real upgrade:
- Reasoning and tool use, natively. The full model takes a configurable reasoning effort (minimal through xhigh). The mini now has a reasoning model behind it, where the previous mini had none at all.
- p95 latency down at least 25% across the Realtime voice models, mostly from better caching.
- Improved alphanumeric recognition (the thing that makes or breaks any agent taking down an order number or a confirmation code) and better silence and noise handling.
The mini’s reasoning capability arrived at the same price as the old non-reasoning mini. That’s the quiet win in this release. If you built a voice agent on gpt-realtime-mini six months ago and haven’t touched the model string since, you’re leaving a free upgrade on the table.
What voice actually costs per minute
Audio token pricing is deliberately hard to reason about, so let’s convert it.
| Model | Audio in / 1M | Audio out / 1M | Cached audio in |
|---|---|---|---|
| gpt-realtime-2.1 | $32 | $64 | $0.40 |
| gpt-realtime-2.1-mini | $10 | $20 | $0.30 |
| gpt-realtime-translate | $0.034/min flat | — | — |
| gpt-realtime-whisper | $0.017/min flat | — | — |
The mini is roughly a third of the flagship on audio, plus $0.60/$2.40 per 1M text tokens, with a 128K context window.
Naively, on the full model, audio input works out to about $0.019/minute and audio output to about $0.077/minute — call it $18/hour for a chatty session. That number is what most vendor comparison posts quote, and it’s wrong in practice for two reasons.
Voice-activity detection. Most of a real call is silence, hold music, or the other person talking. If you’re paying for tokens on all of it, that’s your bug, not OpenAI’s pricing.
Prompt caching. Your system prompt and tool definitions are identical on every single turn. Cached audio input is $0.40/1M against $32/1M fresh — an 80× difference. Aggressive prefix caching alone pulls a typical conversation from ~$18/hour to $6–8/hour.
Do both, and real speech-to-speech traffic lands around $0.04/minute, with $0.04–$0.10/minute being the honest planning range depending on how much text-side reasoning you trigger. And note where the money goes: on a reasoning-heavy agent, the text pass usually costs more than the audio. Optimizing your TTS voice while your agent burns xhigh reasoning on “what are your hours” is fixing the wrong line item.
Direct on the API, or a platform?
The alternative to building on the Realtime API directly is a voice-agent platform, which handles telephony, turn detection, barge-in, call recording, and the fifty other things you didn’t budget for. We went deep on this in our voice agent platform comparison, but here’s how the economics stack against the raw API:
Vapi charges $0.05/minute as an orchestration fee that excludes provider costs. A typical deployment lands at $0.08–$0.15/minute all-in, plus telephony at roughly $0.013/minute per leg.
Retell starts at a flat $0.07+/minute, but a real setup with STT, TTS, LLM, and telephony runs $0.13–$0.31/minute.
ElevenLabs Agents is $0.08–$0.24/minute with everything included in the pipeline — they roughly halved conversational AI pricing in February 2026 after raising $500M at an $11B valuation, which tells you how the competitive pressure in this category is running.
Across the market, per-minute pricing spans about 5× — $0.07 on aggressive self-serve plans up to $0.35 on enterprise tiers with dedicated infra and compliance guarantees.
So: direct API is ~$0.04/minute, platforms are $0.08–$0.31/minute. The platform premium is roughly 2–8×.
Is it worth it? If you’re shipping a prototype, or a single use case, or you don’t have someone who wants to own a WebRTC pipeline at 2am — yes, obviously. The premium buys you the boring 80% of a voice product. If you’re running high call volume with a stable use case, the arithmetic flips hard: at 100,000 minutes/month, the gap between $0.04 and $0.15 is $11,000/month, which funds an engineer to own the stack outright.
The trap is comparing headline rates. Every platform’s advertised number is a platform fee — the LLM, speech-to-text, text-to-speech, and telephony stack on top. When you’re modelling this, insist on an all-in per-minute number or you’ll be off by 2–3×.
What I’d actually do
Shipping this quarter? Build on gpt-realtime-2.1-mini with VAD and prompt caching on from day one, and reserve the full model for the turns that need it. You get reasoning and tool use, ~$0.04/minute economics, and an API that exists today. Don’t wait.
Doing pure translation? Use gpt-realtime-translate at $0.034/minute flat. Don’t route translation through a general voice model and pay conversation rates for it — that’s a specialized endpoint priced specifically to undercut doing it the hard way.
Building a phone-based product? Take a platform. Telephony is a swamp, and $0.10/minute to not personally implement SIP is the cheapest engineering you’ll ever buy. Revisit at volume.
Waiting for GPT-Live’s API? Join the waitlist, but build on the Realtime API anyway. Migration between OpenAI voice endpoints has historically been a model-string change plus some event-handling adjustments, not a rewrite. The cost of building now and migrating later is far lower than the cost of a quarter spent waiting for a launch date that doesn’t exist yet.
The thing worth sitting with: GPT-Live’s real innovation isn’t the full-duplex audio, impressive as it is. It’s the delegation architecture — a fast, shallow model holding the conversation while a slow, deep one does the thinking behind it. That pattern isn’t OpenAI-specific and it isn’t voice-specific. Once the API lands, it’s going to be the default shape of every agent that has to respond faster than it can think, and you can start designing for it now, on the endpoints that already work.
Try this: pull up your current voice agent’s cost breakdown and check what fraction is audio versus text. If you can’t answer that in under a minute, that’s the first thing to fix — before you change a single model.