GPT-Live's one stream vs Gemini 3.8 Live

OpenAI's September 17, 2026 launch post puts GPT-Live at #2 on Artificial Analysis's Speech to Speech Quality Index with a score of 79.4. Google's Gemini 3.8 Live Extended Thinking, announced two days earlier, holds the #1 spot at 82.6 on the same index, according to 9to5Google's report of Google's figures. The gap is 3.2 points. The architectures underneath those scores are not interchangeable, and the pricing is not close either.

GPT-Live is a single speech-to-speech model: one listen-and-speak stream, no separate transcription hop in the middle. Gemini ships two models. The base 3.8 Live is built for fluid dialogue and visual grounding. Extended Thinking is the one that reasons and speaks at the same time, aimed at Gmail Live, Docs Live, and Keep Live. If you only compare the headline index, you miss which product is actually for sale to developers and at what concurrent-session cap.

One WebSocket, two events, no text model in the path

OpenAI's model docs for gpt-live-1 describe a persistent WebSocket. The client streams raw audio in. The model streams raw audio out. Input events and output events share that socket, so the model can be producing speech while it is still receiving speech. Barge-in is a property of the stream, not a second request you fire when the user talks over the assistant.

The launch post is explicit about what is not in the path. There is no speech-to-text model, no text LLM, and no text-to-speech model chained together. OpenAI's line is that the cascaded stack "added latency at every hop and lost information: tone, pacing, interruptions, the thing the user almost said." GPT-Live keeps that signal because the audio never gets flattened into a transcript before the model answers.

Three things run on that same stream, per the launch post:

• Tool use. The model calls tools, narrates what it is doing in audio, and keeps the turn alive instead of going silent while a function returns.
• Interruption handling. Endpointing decides when the user has finished. The model can be cut off mid-sentence without a separate turn detector in front of it.
• Live translation. The same model translates between spoken languages on the stream. The launch post's figure is 98 languages.

Context is 128K. Knowledge cutoff is June 2026. The docs list two snapshots, gpt-live-1 (the alias) and gpt-live-1-2026-09-17. Modalities are text and audio in, text and audio out. Endpoints that accept it: Chat Completions, Responses, Realtime (v1/realtime), Realtime translation, and Realtime transcription. Streaming and function calling are supported. Structured outputs, fine-tuning, and predicted outputs are not.

That last trio is a product decision, not a footnote. If your voice agent depends on JSON schema enforcement or a fine-tune of house style, gpt-live-1 does not offer either. You get function calling and a narration track. You do not get a tunable speech model.

Google's pair: fluid dialogue, then a model that thinks out loud

Google's September 15 announcement, as reported by Abner Li at 9to5Google, calls Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking its "most advanced live dialogue models yet." They follow 3.1 Flash Live, which launched in March 2026. The stated goal is to "make conversing with AI feel more intuitive and intelligent."

The split is the part worth reading carefully.

Gemini 3.8 Live is the conversational model. Google's description, via 9to5Google: "conversational intelligence with fluid dialogue and visual grounding." It processes visual inputs in near real time. Tools and API calls run in the background while the conversation continues. The model is supposed to acknowledge a request and keep chatting while the task finishes. It can detect a language switch and move among 97 supported languages mid-conversation. It takes second place in the Speech Agent Arena, and it will power AI Mode's Search Live experience. Google's framing for this one is scale and cost efficiency, not maximum reasoning.

Gemini 3.8 Live Extended Thinking is the reasoning model. It "reasons and speaks simultaneously." 9to5Google says it uses early verbal cues ("Let me check that…") and live progress narration so a multi-step background task does not produce dead air. The surfaces named for it are Gmail Live (conversational search), Docs Live (draft generation and editing), Keep Live (note creation), plus a rollout into Gemini Live that was starting the day of the announcement.

OpenAI narrates tool use on one model. Google splits the job: a cheaper conversational model that keeps talking while tools run, and a separate extended-thinking model that reasons and speaks on the harder Workspace surfaces. Both companies landed on narration as the way to hide latency. They did not land on the same number of models.

The index, the banking bench, and what neither number proves

Both labs published into the same Artificial Analysis Speech to Speech Quality Index. Put the numbers that are actually in these sources next to each other. Gemini's rows come from the 9to5Google account of Google's announcement. GPT-Live's index score comes from OpenAI's launch post. The other GPT-Live benchmark rows are not in the sources I have, so they stay blank.

Measure	GPT-Live	Gemini 3.8 Live	Gemini 3.8 Live Extended Thinking
Speech to Speech Quality Index (Artificial Analysis)	79.4, #2	Not stated in these sources	82.6, #1 overall
Speech Agent Arena	Not stated	Second place	Not stated
Sierra τ-Voice (agentic task completion)	Not stated	Not stated	68.6%
Sierra τ-Voice-banking	Not stated	Not stated	35.1%
Big Bench Audio	Not stated	Not stated	97.7%
Spoken languages (vendor claim)	98, live translation	97, mid-conversation switch	Not separately stated

A 68.6% on τ-Voice is the lead Google is claiming for agentic voice tasks. The banking slice of that benchmark, τ-Voice-banking, is 35.1%. That is the more useful of the two numbers if you are building anything that moves money or reads an account. Completing about a third of banking voice tasks is not a production voice banker. It is a research score Google chose to publish next to a much friendlier 97.7% on Big Bench Audio, which measures reasoning on audio prompts rather than multi-step tool use against a bank's workflows.

OpenAI did not publish a τ-Voice number for GPT-Live in the launch post. I will not invent one. What OpenAI did publish is the #2 index rank and a list of production behaviors (tool narration, barge-in, 98-language translation) without the agentic completion rate that would let you compare the two on the job Google is advertising.

Big Bench Audio at 97.7% and a Speech to Speech index of 82.6 answer different questions. One says the model can reason about what it heard. The other is a preference-style quality index across speech-to-speech systems. Neither is a latency trace, and neither tells you what happens when the tool call takes four seconds and the user starts talking again. Both vendors describe that case in prose. Neither gives you the milliseconds.

What a session costs, and how many you can hold open

GPT-Live's token prices are on the model page. Text is $5.00 per million input tokens and $20.00 per million output tokens. Cached input is $0.50 per million. Audio is $32.00 per million input tokens and $64.00 per million output tokens. There is no audio cache discount on the price card.

Google's "highly competitive price point" line, as carried by 9to5Google, does not include a rate card in that article. I don't have Gemini 3.8 Live's per-token prices from these sources. Comparing a published $32/$64 audio card to an unpublished "competitive" claim is not a comparison. It is a reminder to go read Google's pricing page before you model a bill.

What OpenAI does specify, and what decides whether a voice product can leave a demo, is concurrency. Rate limits are measured in concurrent sessions, not tokens per minute. The free tier does not include GPT-Live at all.

OpenAI usage tier	Concurrent GPT-Live sessions
Free	Not supported
Tier 1	25
Tier 2	50
Tier 3	200
Tier 4	300
Tier 5	500

Twenty-five simultaneous conversations is the ceiling until your account moves off Tier 1. A support line, a tutoring product, or a translation booth that can have a few dozen people on at once is already past that. Tier 5's 500 is a real production cap, and it is still a cap you have to design for. The docs are explicit that you send gpt-live-1 in the API request. They are not explicit, in the material here, about burst behavior when you hit the session ceiling. Plan on requests failing, not on a queue OpenAI has not described.

Region availability for either model is not in these three sources. If you deploy in a regulated geography, that blank is yours to close with the vendor consoles, not with the launch posts.

Where the two designs actually diverge

Judge them on four things you can verify from the docs and the two announcements: how many models you have to wire, whether reasoning is in the speaking model, what the published price is, and how the language feature is specified.

Wiring. GPT-Live is one model ID on the Realtime socket. Gemini's consumer announcement describes two models pointed at different surfaces. Search Live gets the base model. Gmail, Docs, Keep, and the new Gemini Live rollout get Extended Thinking. A developer comparing API surfaces should not assume those product assignments are the API menu. Google's developer pricing and model IDs are outside this source set.

Reasoning while speaking. Extended Thinking is the model Google says reasons and speaks simultaneously, with filler cues and progress narration. GPT-Live's equivalent claim is narrower and more concrete: on a tool call, it narrates in audio instead of going silent. OpenAI does not, in this launch post, claim a separate "extended thinking" mode that runs alongside speech. One product markets simultaneous reasoning. The other markets an unbroken audio path plus spoken status while tools return.

Price you can put in a spreadsheet. OpenAI's card is complete enough to estimate a voice session if you know your audio token volume. Google's public write-up in this cycle gives you benchmark scores and a relative price adjective. For a finance review this week, GPT-Live is the one you can cost. That is not the same as GPT-Live being cheaper.

Languages. 98 versus 97 is not a decision. The mechanisms differ. GPT-Live offers live translation as a feature of the speech-to-speech model. Gemini 3.8 Live detects a switch and transitions among 97 languages inside an ongoing conversation. Translation and code-switching are different product requirements. A hospital interpreter workflow cares about the first. A bilingual user who starts in English and finishes in Spanish cares about the second. The sources do not say whether GPT-Live code-switches as cleanly as it translates, or whether Gemini's 97-language handoff is a translation product. Don't collapse them.

Visual grounding is the cleanest exclusive. Gemini 3.8 Live takes visual input in near real time while the dialogue continues. GPT-Live's published modalities are text and audio. If the session needs to see a screen, a document camera, or a video frame, the OpenAI model card as written does not cover it.

Fine-tuning is the other exclusive, in the other direction. It is listed as not supported on gpt-live-1, which at least tells you where you stand. Nothing in the 9to5Google piece says whether either Gemini 3.8 Live model can be tuned.

Which session you are actually buying

Buy the OpenAI stream if the product is a voice agent on the API: one socket, tool calls, barge-in, a price card, and a concurrency ladder you can read before you sign the usage tier increase. The #2 index score is what you get in exchange for that operational clarity. You also accept no structured outputs and no fine-tune.

Buy the Google pair, once you have a price and an API ID in hand, if the work looks like the surfaces they named. Extended Thinking is pointed at multi-step Workspace tasks (search a mailbox, draft in a doc, make a note) where the model is supposed to think and talk in the same breath. The base model is pointed at Search Live and at conversations that include a live image. The 82.6 index score and the 68.6% τ-Voice figure are the quality argument. The 35.1% banking number is the argument against treating that quality as competence on high-stakes tool use.

I don't have a head-to-head latency measurement, a Gemini rate card, or an OpenAI τ-Voice score in these sources. Anyone publishing a winner on cost or on agentic voice completion this week is filling those cells from somewhere else.

The figure I keep coming back to is the small one. Extended Thinking leads a benchmark Google wanted published, Big Bench Audio, by enough to look finished (97.7%). On the banking agent benchmark from the same Sierra suite, it completes 35.1%. GPT-Live's launch post never enters that suite. Until both labs report the same agent eval on the same voice stack, the Speech to Speech index is a quality ranking, not a buying guide for anything that has to finish a task.