GPT-Live-1.1 vs Gemini 3.8 Live: Six New Languages Against Ninety-Seven, and Why That Gap Is Not the Whole Story

Count the languages and the fight looks over before it starts. Gemini 3.8 Live, announced by G​oogle DeepMind on September 15, 2026, has carried 97 languages with automatic mid-conversation switching since launch. G​PT-Live-1.1, rolled out by O​penAI on October 15, 2026, adds thirteen new voice options — two each for French, German, Hindi, Japanese, Korean and Mexican Spanish, plus an English voice called Bubbie. Six new languages against ninety-seven is a ratio nobody has to think about.

It is also a comparison between one week of work and one year of coverage, which is why the number that ends the argument is not the number that should decide it. The two releases measure different things: G​oogle's figure describes how much of the world the session can hold, O​penAI's describes which accents were audibly wrong before and are not now. For a buyer, the useful question is not which count is larger. It is which of the two you can actually call, and what happens to the rest of your stack when you do.

Availability is where the comparison breaks

Start there, because it is the point at which these two models stop being alternatives.

G​PT-Live-1.1 is a consumer rollout. It arrived in ChatGPT Voice on October 15, 2026, alongside a second model, G​PT-Live-1.1 Quip, which succeeds G​PT-Live-1 mini and becomes the default for accounts below Plus. O​penAI's announcement is explicit that the API is a later event — "coming soon to the API" — and the documentation agrees: the models page still lists gpt-live-1 and gpt-live-transcribe and nothing newer, and gpt-live-1.1 returns a 404 as a model page. There is no model ID, no rate limit and no price for G​PT-Live-1.1 anywhere a developer can reach.

Gemini 3.8 Live is the opposite case. It is an API model with a published rate — $0.005 per minute of audio input and $0.018 per minute of audio output — and it has been callable since the day it was announced, with the Extended Thinking variant distributed through G​emini Live, Docs, Gmail and Keep for AI subscribers. G​oogle's enterprise availability is still a private preview with no general-availability commitment, which is a real constraint on its own, but it is a different kind of constraint from having no endpoint at all.

So one of these is something you can buy and the other is something you can use. That asymmetry is more consequential than any benchmark below it, and it is the thing most launch-week comparisons of these two models leave out.

What 97 against six actually measures

The gap is real. If your callers speak Polish or Thai, Gemini 3.8 Live is the only one of the two that can take the call, and nothing else in this article matters.

What the number does not measure is quality inside the overlap, and the overlap is where almost all of the demand is.

• What 97 buys — long-tail coverage and the ability to change language mid-sentence without dropping the session. For a support line spanning a dozen European markets, that single feature is the product.

• What six buys — depth in the markets where the previous generation was audibly wrong. O​penAI's claim is about accent naturalness rather than comprehension, and the six languages chosen are among the largest non-English voice markets rather than a random sample. The release notes also describe a wider range of short phrases for signalling that the model is searching, which is the thing users hear as "checking…".

• What neither number tells you — how any of it sounds. There is no public, reproducible accent benchmark and no independent speech-quality score for either model in either language set. Both figures are vendor counts of supported locale options, and both vendors publish them without a method attached.

The practical consequence is that language count is a screening filter, not a decision. If your callers speak one of the seven languages above, the count is irrelevant and accent quality becomes a listening test you run yourself. If they do not, the decision has already been made for you.

The independent board, and the variant problem

Where independent measurement exists, it exists for the previous generation on the O​penAI side. On the Artificial Analysis Speech to Speech Index, read from the board as it stood on October 16, 2026:

• G​rok Voice 3.0 — 86.9

• G​PT-Live-1 (Astra backend, medium effort) — 83.2

• Gemini 3.8 Live Extended Thinking (High) — 82.6

• G​rok Voice 2.0 — 81.3

• G​PT-Live-1 (Sol backend, low effort) — 80.1

• Gemini 3.8 Live — 76.0

• G​PT-Realtime-2.1 High — 75.2

• G​rok Voice 1.0 — 73.7

• G​PT-Realtime-2 (High) — 73.6

• Gemini 3.1 Flash Live High — 72.0

Two things fall out of that ordering. First, the model named second in this article's title sits sixth while a variant of the other one sits third — the 82.6 that gets quoted belongs to Gemini 3.8 Live Extended Thinking, a different configuration with a narrower rollout. Second, there is no G​PT-Live-1.1 row anywhere on the board, so the successor is being represented here by the number its predecessor earned.

Read against the standard configuration — Gemini 3.8 Live's 76.0 against G​PT-Live-1's 83.2 — the O​penAI family leads by 7.2 points on the predecessor's evidence. Read against the Extended Thinking variant it is a 0.6-point gap the other way, which is inside the movement you expect from a board that gets re-run as models are added. Neither reading is a measurement of G​PT-Live-1.1, and it is worth being blunt about that: the successor's own evidence is a system-card statement that it performs at parity with G​PT-Live-1 on production prompts, which is a claim about not regressing, not a score.

Where architecture still separates them

The full-duplex difference has not gone away. It has just stopped being decisive.

G​PT-Live-1.1 inherits the structure of G​PT-Live-1: audio in and audio out processed continuously inside one model, no speech-to-text → language model → text-to-speech chain, backchannels like "mhmm" produced natively, and interruption handled by deciding several times a second whether to keep talking or yield. That remains the thing turn-based systems get wrong, and Gemini 3.8 Live is still turn-based.

What Gemini 3.8 Live does instead is keep the session productive while a turn is in flight: near real-time visual input processing, background tool and API execution that acknowledges a request and keeps the conversation alive while the call completes, and an Extended Thinking variant that narrates its progress out loud. G​oogle's answer to the silence problem is to fill the silence with work rather than to remove the turn boundary. Both approaches stop a caller from hanging up; they fail differently, and turn-based models can still misjudge a pause.

Visual input is the concrete cost of that trade. Camera and screen sharing have been part of Gemini 3.8 Live's surface since launch, and O​penAI's own documentation for the Live family lists image and video input as explicitly unsupported — a limitation that carries forward to G​PT-Live-1.1, since the consumer help pages for the new version repeat that Live does not support video or screen sharing. If your agent needs to look at something, this decides the comparison before any benchmark does.

The cost comparison, with the asterisk attached

On published front-end rates, G​oogle is cheaper by a wide margin. Gemini 3.8 Live bills $0.018 per minute of audio output against G​PT-Live-1's $0.05 per voice minute, and that $0.05 is the only per-minute figure O​penAI has published for the Live family — it belongs to the older model, and no rate has been announced for G​PT-Live-1.1 because no endpoint exists to attach one to.

The asterisk is that neither number is the bill. G​PT-Live-1's $0.05 covers the voice front-end only. The reasoning happens on a delegated text model, and O​penAI's Live documentation states plainly that backend Responses calls are billed at the configured model's normal pricing — so five cents a minute of voice time becomes three dollars an hour of voice time before a single reasoning token is counted. Gemini 3.8 Live has the identical shape: background tool execution is text-model work, billed at text-model rates that G​oogle has not published for this product.

That is why any per-hour cost figure you see for either model is a partial number wearing a total's clothes. An hour of G​PT-Live-1 voice time is $3.00 of front-end before the backend is metered at all, and the all-in figure depends entirely on which model you delegate to and how often it gets called. For Gemini 3.8 Live the front end is cheaper per minute and the backend is undisclosed. Comparing the two headline rates tells you about the cheap half of each invoice.

The layer both of them hand work to

Both systems delegate, and that is the fact that makes this matchup less of a lock-in decision than the language count suggests. G​PT-Live-1.1 hands hard queries to a backend text model by design; Gemini 3.8 Live's background tool execution resolves into ordinary model calls. In both cases the voice layer is a thin front end over a reasoning model that the vendor treats as separately configured and separately metered. Cost variance lives there, quality differences between configurations show up there, and switching is cheap there because it does not touch the voice integration.

That is the half of a voice stack where a router earns its place, so it is worth being precise about the boundary. On OrcaRouter the reasoning layer is more than 200 models behind one key at provider list price with no markup, which means an upstream price cut reaches our side the same day rather than at the next contract renewal — a property that matters more than usual in a two-meter billing structure where the backend is the meter that moves. The routing DSL lets you name the delegation target explicitly and swap it on live traffic, and automatic failover keeps a session alive when a backend errors or times out mid-call, which is a failure mode voice has and text does not. One clarification, because it would be easy to imply otherwise: we do not host the G​PT-Live-1.1 or Gemini 3.8 Live voice endpoints. Those come from the vendors' own platforms, and in G​PT-Live-1.1's case there is no public endpoint to host. The text models these systems hand work to generally are on the router.

How to choose

Take Gemini 3.8 Live if you need to ship now, if your languages extend past the seven named above, or if the model has to see something. Accept that you are buying the standard configuration that scores 76.0 on the composite index, that the 82.6 usually quoted belongs to the Extended Thinking variant, that enterprise access is a private preview, and that the all-in cost per hour is unpublished on your side of the invoice too.

Take G​PT-Live-1.1 if conversation quality is the product and your languages sit inside its set — and if you can wait. Full-duplex turn-taking is still the naturalness benchmark, the version bump is a real improvement over G​PT-Live-1, and the accent work targets exactly the languages that were worst before. It is also, today, not something you can call from code.

Take G​PT-Live-1 if you need voice from O​penAI and you need it this month, because it is the only member of that family with a model ID, a documented endpoint and a published rate.

The date to watch is the one where gpt-live-1.1 gets a model page. That is when this stops being a comparison between a shipping product and an announcement, and when the accent claim, the language count and the price all become testable in the same afternoon.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.