Two full-duplex voice models shipped in five days this month, and on the benchmark everyone quotes for how natural they sound, you cannot tell them apart. OpenAI put GPT-Live-1 in the API on September 10. Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 15. On Artificial Analysis’ Full Duplex Bench, which scores conversational dynamics like turn-taking and interruption handling, GPT-Live-1 lands at 97.3% and Gemini 3.8 Live at 96.1%, with StepAudio 3 Realtime (98.9%) and Qwen Audio 3.0 Realtime Plus (98.4%) sitting above both. Four models within three points of each other. Voice quality is no longer the axis you choose on.
That sounds like a small point. It is actually the whole buying decision, because once naturalness converges, what separates these products is where the reasoning lives, what latency you can hold, what a minute costs at concurrency, and whether you can see what the agent said. Those are operations questions, not demo questions. A voice agent that scores 97% on a naturalness bench and blows a 400ms latency budget on every call is a worse product than a plainer one that stays under 250ms. I have spent enough years running telephony and contact-center infrastructure to know that the thing that kills a voice deployment is almost never how the synthesized voice sounds.
Where each model puts the brain
The genuinely interesting split in this month’s releases is architectural, and OpenAI and Google landed on opposite sides of it.
GPT-Live-1 is a full-duplex voice layer and nothing more. It handles incoming and outgoing audio in one model, replacing the old chain of speech recognition, a text model, and speech synthesis, but it does not do the thinking. You attach a separate reasoning model behind it, and OpenAI names GPT-6 Astra, Luna, and Terra as options alongside third-party models. The voice layer is priced at $0.05 per minute, billed per second, with that backend model and any tools billed separately. So the design keeps your model choice open: the voice is fixed, the brain is swappable, and you can put a cheap model behind casual chit-chat and route the hard turns to Astra the same way you would tier any other workload.
Gemini 3.8 Live took the other bet. It fuses the reasoning into the speech model, so it hears prosody, tone, and hesitation directly rather than through a transcript that has already discarded them, and the Extended Thinking variant runs background reasoning in parallel while it keeps streaming audio. It scored 68.6% on the τ-Voice agentic benchmark, 97.7% on Big Bench Audio, and topped Artificial Analysis’ Speech-to-Speech Quality Index at 82.6. The pricing is roughly $0.005 per minute of audio input and $0.018 per minute of output. Function calls run asynchronously, so the agent can query a database or update a system while it is still talking.
The third architecture is the one neither company shipped this month but plenty of teams still run: the cascaded pipeline of speech-to-text, then a language model, then text-to-speech, stitched together with something like ElevenLabs in the synthesis slot. It is worth understanding why anyone still builds this way, because the reasons are about control, not nostalgia. A cascade transcribes before it reasons, so you get a text log of every call for free, which matters for compliance and for evaluation. Each stage is independently swappable. The cost is latency: a cascade’s total delay is the sum of three stages, and it loses the tone and hesitation that a native model hears directly.
So there are really three products in front of you: a modular voice layer with a brain you rent (GPT-Live-1), a fused model that reasons in audio (Gemini 3.8 Live), and a pipeline you assemble yourself (the cascade). They fail in different ways, and they bill in different ways.
The number that decides it is latency, not naturalness
Humans hand conversational turns back and forth in roughly 200 milliseconds, and the same benchmark work notes that anything above 800ms feels noticeably delayed while anything above 1,500ms makes callers feel the conversation is broken. The ITU-T G.114 telephony standard has recommended no more than 150ms of one-way transmission delay for good interactive quality for decades. Voice agents live inside that budget or they annoy people.
That budget is tighter than the marketing suggests. The practical 2026 production targets, per FutureAGI’s implementation work, are a p50 under 250ms on an optimized stack, under 400ms on a standard cloud stack, and a p95 under 800ms either way, with barge-in (the agent shutting up the instant the caller starts talking) needing to fire in under 150ms. Miss the barge-in window and the agent talks over people; miss end-of-turn detection, which typically waits 800 to 1,200 milliseconds of silence plus a semantic check that the sentence looks finished, and it either interrupts or leaves dead air.
The launch pages do not make this easy to compare. OpenAI published no latency number for GPT-Live-1 at all. Gemini 3.8 Live posts about 1.18 seconds to first audio, rising to 1.35 seconds for the Extended Thinking variant, per orcarouter’s teardown, though time-to-first-audio is a different measurement than round-trip turn latency. And the model is only one term in the sum. Telnyx’s own testing puts carrier-leg round trip at 71ms p50, and full voice-to-voice latency on hosted agent platforms measured out to 680ms median for Retell and 720ms for Vapi across 500 calls. Your network, your telephony carrier, and your tool calls all spend from the same 800ms ceiling the model does. Benchmark the whole path on your own traffic, not the model in isolation.
Cost stops being a rounding error at concurrency
At one call, none of this matters. At a contact center running thousands of concurrent minutes, it is the budget line that gets you fired.
Take GPT-Live-1’s voice layer alone at $0.05 per minute. That is $3,000 for 1,000 hours of talk time before the backend reasoning model or any tool call is billed. Put Astra behind it for the thinking and you are paying that voice-layer rate plus Astra’s per-token rate plus whatever your tools cost, all metered per minute of a live human conversation that you do not control the length of. Gemini 3.8 Live’s fused pricing works out to roughly two cents a minute for a talk-heavy call, which is cheaper per minute but folds the reasoning into that number so you cannot dial the brain down independently. Neither is expensive at demo scale. Both compound fast when a queue of callers is holding lines open, and the same tiering discipline that governs text workloads applies here with a stopwatch attached. Model the cost per completed call at your real concurrency, not the sticker rate per minute.
A voice agent that can act is an untyped input channel
Here is the part that should worry anyone deploying this into production, and it is the part the launch posts skip. A voice agent on a phone line takes untrusted audio from anyone who dials in and, increasingly, can act on it: look up an account, change a booking, move money. That is the same lethal-trifecta shape that makes AI browsers a security decision, only the injection now arrives as speech and there is no address bar to inspect. Prompt injection over audio is a live research problem, and the mitigations are not solved.
Three controls are non-negotiable before a voice agent goes near a phone number. First, recording and consent: a live call is regulated, and two-party-consent jurisdictions require disclosure before you capture audio. This is where the cascade’s free transcript is a genuine advantage, and where a native model forces you to build logging you would otherwise get for nothing. Second, treat the agent’s action permissions like a privileged service account, scoped to the narrowest set of operations the call type actually needs, with a human confirmation gate above any consequential action, the same way you would handle an agent delegating to tools in any other production system. Third, keep the transcript, because when a caller claims the agent authorized a refund, “we do not log audio” is not an answer you want to give.
How I would actually pick
If you are in a regulated environment or you will ever need to prove what was said, the transcript question decides it before latency does: run a cascade, or commit up front to logging every native call. If naturalness and the lowest possible latency are the product, and you operate the surrounding stack, the fused native model (Gemini 3.8 Live, and its Extended Thinking variant for tasks that need real reasoning mid-call) is the strongest bet. If you already have a text-model stack you trust and you want to keep choosing the brain per call type, GPT-Live-1’s split design is built for exactly that, and it drops onto telephony without you assembling a pipeline.
Whichever way you lean, the decision is not “which one sounds most human,” because on that question the field has converged and the leaderboard will mislead you. Wire up two candidates against a dozen of your own real call types, measure round-trip latency and cost per completed call at your actual concurrency, and pick on those numbers. The voice was the easy part. It always was.