Gemini 3.8 Live vs ElevenLabs: A Model Against a Pipeline

Open ElevenLabs' documentation for its agent platform and there is a Gemini model sitting in the middle of it. ElevenLabs Agents is a cascade — a speech recognition front end, a language model, and a text-to-speech back end — and the language model in the published stack is a Gemini one. That is the right place to start a Gemini 3.8 Live vs ElevenLabs comparison, because the two things being compared are not the same kind of object. Gemini 3.8 Live is the speech-to-speech model, shipped on the Live API on September 15, 2026, priced per minute of audio. ElevenLabs Eleven v3 is the flagship synthesis model of a voice platform that also sells ElevenLabs Agents, and it is priced per character, not per conversation. One is a component you can rent. The other is a system that rents components — including, in at least one published configuration, a Gemini model's.

That asymmetry settles most of the questions people actually bring to this matchup. If you are asking which one to call, the honest answer is that they are not substitutes: one is a model with a rate card and no telephony, the other is a product with telephony, concurrency limits, and three meters running at once. Which one is cheaper, better, or faster depends entirely on which of those two shapes your application already is.

One model, or three models in sequence

Gemini 3.8 Live is a native speech-to-speech system. Audio goes in, audio comes out, and the reasoning happens in the same forward pass that produces the speech — one model, one latency budget, one bill. Google shipped it in two variants on September 15, 2026: gemini-3.8-live for conversational work at scale, and gemini-3.8-live-extended-thinking for the same interface with multi-step reasoning behind it. Both handle 97 languages with mid-conversation switching, both process visual input in near real time, both run tool and API calls in the background while the conversation continues, and both watermark every second of generated audio with SynthID. The published rate is $0.005 per minute of audio input and $0.018 per minute of audio output.

ElevenLabs Agents is not that. It is a pipeline, and the pipeline is the product. A real-time speech recognition model transcribes the caller, a language model decides what to say, and a synthesis model — Eleven Flash v2 in the configuration ElevenLabs documents — speaks the answer. Each stage is metered separately, each stage can be swapped, and the whole thing sits behind a managed layer that handles telephony, turn detection, and concurrency. ElevenLabs Eleven v3, the company's headline synthesis model, is a different product again: a text-to-speech model priced at $100 per million characters, sold for narration, dubbing, and voice design rather than for holding a conversation.

So the first thing to get right is that "Gemini 3.8 Live vs ElevenLabs Eleven v3" is not a head-to-head between two speech models. Eleven v3 does not take audio input and does not hold a conversation. The comparison that means something is Gemini 3.8 Live against ElevenLabs Agents, with Eleven v3 in the room only as the synthesis quality ceiling the platform is built around.

Where each one actually sits on a board

There is no leaderboard that ranks these two against each other, and the reason is instructive: they keep showing up on different boards measuring different things.

On Artificial Analysis's Speech to Speech Index — which blends speech reasoning, agentic task completion, arena preference, and task success into one number — Gemini 3.8 Live scores 76.0, with the Extended Thinking variant at 82.6 in first place. Those are measurements Artificial Analysis runs under its own harness, not figures Google published, though Google's own announcement quotes numbers from the same components.

ElevenLabs does not appear on that board at all, because it does not ship a speech-to-speech model to measure. Where it does appear is the Speech Agent Arena, which evaluates assembled agents rather than models, and the picture there is genuinely mixed: ElevenLabs Agents has been measured at 937 Elo with a 90.5% task success rate, against 1,046 Elo and 74.6% task success for an agent built on the previous Gemini Live generation, with the Gemini agent reaching first audio in 0.96 seconds. Read that pair carefully. The Gemini-backed agent wins on preference and speed; the ElevenLabs agent wins on getting the task done. That is a cascade beating a native model on reliability, which is not the result the architecture diagrams would predict.

Two caveats on those figures, both of which matter. The Gemini side of that comparison used the early-2026 Gemini Live generation, not 3.8 Live, so it is not a reading on the model this article is about. And the third board in play — Artificial Analysis's Provider Voices speech arena, which ranks synthesis models — puts ElevenLabs Eleven v3 at 1,177 Elo and the conversational variant, ElevenLabs v3 Conversational, at 1,219 Elo, against 1,283 for the leader, Cartesia Sonic 3.6. That board measures how good the voice sounds, not how well the agent performs, and mixing the two is how people end up quoting a synthesis score as evidence about a conversational system.

Spec contrast

• Architecture — Gemini 3.8 Live is one native speech-to-speech model vs ElevenLabs Agents is a cascade of speech recognition, a language model, and synthesis

• Input and output — Gemini 3.8 Live takes audio and returns audio in a single pass vs ElevenLabs takes audio through Scribe v2 Realtime and returns it through Eleven Flash v2, with a text transcript in between

• What you can swap — Gemini 3.8 Live nothing, the model is the model vs ElevenLabs every stage independently, including the language model, which in the documented stack is a Google one

• Languages — Gemini 3.8 Live 97 with mid-conversation switching vs ElevenLabs Eleven v3 74

• Audio pricing — Gemini 3.8 Live $0.005 per minute in and $0.018 per minute out vs ElevenLabs priced on agent minutes with language-model tokens metered separately on top

• Telephony — Gemini 3.8 Live none first-party, you bring your own carrier and session management vs ElevenLabs managed, first-party

• Concurrency — Gemini 3.8 Live whatever your Live API quota allows vs ElevenLabs sold in concurrency tiers

• Agent tooling — Gemini 3.8 Live background tool and API execution inside the model vs ElevenLabs a managed agent layer with turn detection, workflows, and a voice library

• Open artifact — neither. Both are closed, cloud-only, and proprietary

• Latency — Gemini 3.8 Live 1.18 seconds to first audio for the standard model and 1.35 for Extended Thinking, per Artificial Analysis vs ElevenLabs not published as a single number, because a cascade's latency is the sum of three stages

The last row is the one that explains the architecture choice better than any of the others. A native model has one latency budget. A cascade has three, and the reason teams still pick the cascade is everything above it: the transcript you can log, the stage you can swap, and the phone number that just works.

What a minute costs, both ways

Gemini 3.8 Live's arithmetic is unusually easy because there is only one meter. A minute of conversation is roughly a minute of caller audio plus a minute of model audio: $0.005 plus $0.018, so about 2.3 cents a minute at the published rate. Artificial Analysis's cost-per-hour column, which normalises the cost of completing a fixed audio-reasoning set to an hourly figure, puts the standard model at $0.84 per hour of input audio — the cheapest paid rate on that board — and the Extended Thinking variant at $3.50.

ElevenLabs' arithmetic is harder, and that is the point rather than a criticism. An agent minute is not one price: there is the platform charge for the agent minute itself, the language-model tokens consumed on the middle leg, and the carrier minutes underneath — three meters, three vendors in the worst case, and a bill that moves when any one of them moves. The synthesis side is more legible: ElevenLabs Eleven v3 at $100 per million characters is roughly $8 an hour of continuous narration at ordinary speaking rates, against about $2.70 for the cheapest open-weights competitor on the same arena. ElevenLabs charges a premium for the voice, and on that board the premium is real — it sits fourth overall.

What the two cost structures mean in practice is that Gemini 3.8 Live is cheaper per minute of conversation and far cheaper per hour of audio, while ElevenLabs is more expensive and far more predictable to operate. A team with no ML engineers and a phone number to stand up will spend less on ElevenLabs in the first month, because the alternative is building what ElevenLabs already built. A team already running its own session management and its own carrier will spend less on the raw model, because they are not paying for a pipeline they already have.

What ElevenLabs is genuinely better at

It would be easy to read the price gap as a verdict. It is not, and the reasons are the ones a rate card never shows.

• Telephony. ElevenLabs sells phone numbers, SIP trunking, and the concurrency to answer calls. Google sells a model that speaks. If your product is a phone line, the gap between those two sentences is months of engineering.

• Voice identity. The Voice Library, the cloning pipeline, and the dubbing studio are a mature product surface. Gemini 3.8 Live gives you a model; it does not give you a catalogue of licensed voices or a workflow for localising an existing recording into nine languages.

• A transcript by default. Because a cascade transcribes before it reasons, you get text logs of every call for free — useful for compliance, for evaluation, and for the unglamorous work of finding out why the agent got something wrong. A native speech-to-speech model has no such artefact unless you build one.

• Swappable parts. When a better language model ships, an ElevenLabs cascade can adopt it without retraining anything. That is exactly why the documented stack has a Google model in the middle of it — and it is the single strongest argument for the cascade architecture.

What the raw model buys you

The counter-argument is not about price. It is about what a single forward pass can do that three stages cannot.

Gemini 3.8 Live hears prosody, tone, and hesitation directly rather than through a transcript that has already thrown them away, which is why it can switch languages mid-sentence and why its conversational-dynamics score — 96.1% for the standard model, per Artificial Analysis — is the one component where it beats its own reasoning-heavy sibling. It also runs tool and API calls in the background while it keeps talking, so a lookup does not produce silence. And the Extended Thinking variant is a genuinely different capability rather than a bigger version of the same one: it resolves 68.6% of τ-Voice customer-service scenarios against 30.1% for the standard model, a 38-point gap that is the largest single number in the release and the reason Google split the line in two.

If your agent's job is to finish a multi-step task rather than to hold a pleasant conversation, that 38-point gap is the entire decision, and it costs $3.50 an hour instead of $0.84 — still below every model it beats on that board.

The middle leg is the one you can actually move

Here is the seam worth naming, because it is where the architecture question turns into a procurement question. In a cascade, the middle leg — the part that decides what to say, calls the tools, and does the reasoning — is ordinary text inference. It is also the leg with the most cost variance, the most lock-in, and the least reason to be tied to any single vendor. The stack ElevenLabs documents happens to use a Gemini model there. It does not have to.

OrcaRouter covers that leg and not the outer two. We do not host the Gemini 3.8 Live voice endpoints — those come from Google's own Live API — and we do not host ElevenLabs, whose synthesis and agent layers are its own product. What we carry is the text layer underneath: 200+ models behind a single key at provider list price with no markup, so a price cut from an upstream lab reaches your bill the same day rather than at the next renewal. For a voice stack specifically, three of those properties earn their keep. Automatic failover keeps a call alive when a backend errors or times out mid-sentence, which is a failure a caller notices immediately. The routing DSL composes several models into one call, so a cheap model can handle the acknowledgement turns and a stronger one can take the transaction. And model fusion lets a panel answer together on the turns where a single model's judgment is not good enough to be trusted alone. Changing which model sits in the middle of a cascade is a configuration change, not a rebuild — which is the whole reason the cascade architecture exists.

Which one to pick

Pick ElevenLabs if you are shipping a phone line and you do not want to build a voice stack. You are buying telephony, a voice catalogue, transcripts, concurrency, and a support contract, and the per-minute premium is what those cost. You are also buying the ability to change the model in the middle without changing anything else, which is worth more than it sounds.

Pick Gemini 3.8 Live if the voice is a feature of something you already operate. You get the cheapest competent speech-to-speech rate anyone currently publishes, 97 languages with mid-conversation switching, tool execution that runs while the model keeps talking, and — if your agent has to complete tasks rather than just converse — the 38-point agentic gap that the Extended Thinking variant buys for about four times the hourly audio cost. You supply the carrier, the session lifecycle, and the logs.

What to watch: whether ElevenLabs publishes a single latency figure for a managed agent minute, because that is the number that would make this comparison quantitative instead of structural; and whether the Speech Agent Arena result holds up when an agent built on 3.8 Live is measured on it, because a cascade currently beating a native model on task success is the most interesting thing in this matchup and the least explained.