A generated title card for Realtime 1 vs Gemini 3.8 Live contrasting a 1074 arena Elo and 93.2% task success rate against a 30.1% agentic task-completion score, plus an OrcaRouter logo.
Guides & Insights

Realtime 1 vs Gemini 3.8 Live: The Arena Winner Is the One That Can't Finish the Task

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

G​emini 3.8 Live has the best arena Elo of any voice model on Artificial Analysis' speech-to-speech board — 1074, ahead of GPT-Live-1's 1037 and well clear of Eleven Realtime Max's 990. It also has the highest task success rate in that comparison set, 93.2%. And it scores 30.1% on agentic task completion, which is the weakest figure of the four. Those three numbers describe the same model, and understanding why they coexist is the whole question when you put it next to MAI Realtime 1, Microsoft's new speech-to-speech model announced on 17 November 2026, which is pitched at exactly the thing G​emini 3.8 Live is worst at.

One of these models is a conversationalist. The other is being sold as an operator. Neither has published enough to prove which is which on your workload, but they fail in opposite directions, and that is more useful than a scoreboard where everything is close.

Two different bets about what a voice model is for

G​oogle's framing for G​emini 3.8 Live is narrow and confident: it is the default Live API model for "most low-latency voice agent experiences without reasoning delays," with a sibling — G​emini 3.8 Live Extended Thinking — for cases that need "higher background reasoning" during the interaction. That split tells you what G​oogle thinks the base model is not for. Real thinking lives in the sibling.

Microsoft's framing for MAI Realtime 1 is the mirror image. It is "speech-to-speech with natural conversational flow and precise steerability," and the architecture is explicitly two-tier: a fast system keeps dialogue real time, and harder tasks are "routed to a stronger reasoning engine without awkward pauses." Microsoft is not selling a better talker. It is selling a dispatcher that talks well enough to keep the line alive while something else does the work.

• G​emini 3.8 Live — one model, deliberately fast, with reasoning available as a separate, heavier sibling you choose per deployment.

• MAI Realtime 1 — one product with two engines behind it, escalating per turn rather than per deployment.

• The practical difference — G​oogle asks you to pick a setting at integration time. Microsoft asks you to trust a router inside the model and pay its reasoning tier by the token.

What the independent numbers say about G​emini 3.8 Live

This is the side of the matchup with actual measurements, and they are good:

• Speech-to-speech index — 74.5, against 82.6 for the Extended Thinking variant. The index penalises it for the agentic column, not for the conversation.

• Speech reasoning (Big Bench Audio) — 91.7%. Strong, and 97.7% on Extended Thinking.

• Conversational dynamics — 96.1%. This is where the arena Elo comes from: it handles the human rhythm of a call about as well as anything measured.

• Arena Elo — 1074, the highest in the comparison set. Human preference, in effect, and it is not close to the bottom of that list.

• Task success rate — 93.2%, also the highest in the set.

• Agentic task completion (τ-Voice) — 30.1%. The outlier, and the reason the composite index sits below the extended-thinking version.

• Time to first audio — 1.18 seconds. Not the fastest — Eleven Realtime Max does it in 0.57 — but fast enough that the "no reasoning delays" claim is credible.

• Blended cost — about $0.84 per hour on Artificial Analysis' normalisation. G​oogle's own rate card reads $0.005 per minute of audio input and $0.018 per minute of audio output, which on a continuous two-way call is closer to $1.38 an hour; the gap is a measurement convention, not a contradiction, and neither figure should be quoted as the other.

Read the agentic number against the task-success number and something important shows up. A 93.2% task success rate next to 30.1% agentic completion means the model is very good at the tasks the harness considers conversational and poor at the ones that require it to plan across steps and use tools. Those are different instruments, and the disagreement is the finding. G​emini 3.8 Live will hold a better conversation than MAI Realtime 1 probably will. Whether it will book the appointment at the end of that conversation is a separate question, and on the only public evidence available it is the weaker half.

A screenshot of Artificial Analysis's Speech to Speech AI Model & Provider Leaderboard page captured 18 November 2026, showing the page heading, the methodology summary covering reasoning quality, conversational dynamics, generation time and price, the Speech to Speech Index / Speed / Cost per Hour chart panels, and the Related Links rail; the individual model rows in the charts are not legible at this capture size.

What the Microsoft numbers say, which is one number

MAI Realtime 1 has no independent scores. Artificial Analysis does not carry the model as of 18 November. What Microsoft publishes itself is a single accuracy figure on the model page — 98.8% on Big Bench Audio, labelled there as accuracy on spoken reasoning questions from that benchmark — plus a P50 time-to-first-audio card and a "P50 Latency on BigBench Audio" caption. No τ-Voice figure, no supported-language list, no distribution behind the percentile, no methodology note beside the accuracy number. Compare that to G​emini 3.8 Live's 91.7% on the same benchmark as measured by Artificial Analysis, and you are comparing a vendor's own evaluation against a third party's: the two numbers are not the same kind of object, and treating 98.8% as a seven-point win would be exactly the laundering this piece is trying to avoid. The company's broader claim is that the model sits on the performance-price Pareto frontier for Big Bench Audio and τ-Voice, and that it beats G​PT Realtime 2.1 on quality, speed and price — a comparison against O​penAI's per-token reasoning voice model, not against G​emini 3.8 Live.

So any direct comparison between these two models today is a comparison between a third-party measured profile and a vendor's own headline figure. That is not a reason to skip the matchup. It is a reason to be exact about what each side has actually committed to.

What Microsoft has committed to, concretely: $0.01 per audio minute, reasoning at $1.00 per million input tokens and $5.00 per million output tokens, an O​penAI Realtime API-compatible interface, a limited-access Foundry preview that is unbilled until public preview, a two-minute non-commercial demo in MAI Labs restricted to adults, a vendor-reported 98.8% on Big Bench Audio with a P50 latency figure and no distribution behind it, named deployments coming in Dynamics 365 Contact Center and Dragon Copilot's nurse-training simulations, and early testing at Agoda and Reservations.ai.

What it has not committed to: languages, τ-Voice, data residency, retention, or a GA date.

A screenshot of Microsoft's MAI Realtime 1 model page captured 18 November 2026, showing the 98.8% Big Bench Audio accuracy figure, a 'Start a conversation' button, $0.01 per audio minute pricing, $5 per 1M output tokens, and a P50 time-to-first-audio card labelled 'P50 Latency on BigBench Audio'.

The cost comparison, done properly

On the voice line alone the gap is not subtle. A minute of G​emini 3.8 Live audio, input and output combined, is $0.023 on G​oogle's card. A minute of MAI Realtime 1 is $0.01 before reasoning. Roughly 2.3x, and unlike the ElevenLabs matchup there is no compliance or platform fee on the G​oogle side to widen it.

But the interesting cost difference is not the rate. It is where the reasoning bill lands. G​emini 3.8 Live's answer to a hard turn is a different model you deploy — G​emini 3.8 Live Extended Thinking, at $3.50 per hour on Artificial Analysis' basis and 82.6 on the index. That is a deployment decision, priced per hour of audio whether or not any given turn needed it. MAI Realtime 1's answer is a per-turn escalation priced per token, which is cheaper when few turns are hard and more expensive when most are.

• Mostly easy turns, occasional escalation — MAI Realtime 1's per-token model wins, and by a lot.

• Mostly hard turns — G​emini 3.8 Live Extended Thinking's flat $3.50 an hour is the more predictable instrument, and predictability is worth something when you are forecasting a contact-centre budget.

• Mixed traffic with no clear split — you are choosing between a per-hour estimate you can defend and a per-token estimate you have to measure. Those are different procurement conversations.

Worth noting that G​oogle's free tier covers G​emini 3.8 Live at no charge for input and output, rate-limited. For a prototype that is a real advantage over a registration form and a waitlist.

Languages, which decides more matchups than benchmarks do

G​emini 3.8 Live inherits the Live API's multilingual behaviour and G​oogle's documentation treats non-English dialogue as ordinary use. MAI Realtime 1's launch material does not list a single supported language. Microsoft's voice lineage runs through Dragon, which is a clinical dictation product with strong English deployment, and the Transcribe 2 announcement talks about halving multilingual error rates — but that is the transcription model, not the realtime one, and it is not a language list.

If your agent serves one English-speaking market, this is a footnote. If it serves three, MAI Realtime 1 is not yet a candidate you can evaluate, because the preview will not tell you what it does in the languages you need.

How to run the comparison before either vendor settles it

Both models speak the O​penAI Realtime API — G​oogle's Live API through its own surface, Microsoft's explicitly by design. The G​emini side is closely related but not identical in session semantics, so budget for an adapter rather than a base-URL swap; the MAI Realtime 1 side is closer to a drop-in by Microsoft's own description.

The evaluation that decides this matchup is not a benchmark run. It is a task-completion run: give each model the same twenty real requests that require a tool call and a second step, and count how many finish. G​emini 3.8 Live's 30.1% agentic figure sets your prior — it should finish roughly six. If MAI Realtime 1 lands near that, the price advantage buys you nothing on this workload. If it lands near the 87% that Eleven Realtime Max manages on the same harness, Microsoft's two-tier design is doing what it claims and the matchup is over.

That evaluation needs the reasoning layer behind the voice model to be instrumented, because on the Microsoft side that is where the second step actually happens. OrcaRouter is the place that layer plugs in: 200+ models behind one key at provider list price with 0% markup, automatic failover so a reasoning call that stalls does not become four seconds of silence on a live call, and a routing DSL that lets the escalation path be a rule rather than code you maintain in the voice handler. MAI Realtime 1 itself is not in our catalogue — it is reachable through MAI Labs and the Foundry limited-access preview — so what we route is everything it hands off to, plus G​emini 3.8 Live's siblings if you go the G​oogle way.

What the preview does to the procurement timeline

Put the two availability routes side by side and the asymmetry stops being about the models and starts being about how long it takes to know anything. G​emini 3.8 Live is generally available on G​oogle's Live API today, with a free tier covering input and output at a rate limit, which means a two-day spike can produce a task-completion number before the end of the week. MAI Realtime 1 is a registration form, a waitlist, a two-minute non-commercial demo restricted to adults, and a preview Microsoft has said it will not bill until public preview begins.

Those are different kinds of friction and they favour different teams. If you need an answer now and the answer has to be defensible to someone who signs the invoice, G​emini 3.8 Live gives you a measured profile with a public harness behind it and a published rate you can argue about. If you have engineering capacity to spare and no deadline this quarter, the Microsoft preview is the cheapest useful experiment in the category — a cent a minute that costs nothing today, against a model no third party has measured, on an interface you probably already speak.

• Prototype this week, decide next quarter — either works, and the free tiers on both sides mean the experiment is almost free.

• Prototype this week, ship next month — G​emini 3.8 Live, because MAI Realtime 1 has no GA date and no way to forecast one.

• Data cannot leave your tenancy or your region — neither of these is the answer on published terms; G​oogle has regional controls on Live API, Microsoft has said nothing about residency for Realtime 1.

A generated two-column scoreboard titled 'MAI Realtime 1 vs Gemini 3.8 Live — the scoreboard'. Left column MAI Realtime 1: arena Elo not scored, task success not published, agentic not published, reasoning 98.8% vendor, cost $0.01 plus tokens. Right column Gemini 3.8 Live: arena Elo 1074, task success 93.2%, agentic 30.1%, reasoning 91.7%, cost $0.84 per hour. Footer reads 'Gemini figures per Artificial Analysis; the 98.8% is Microsoft own, unaudited.'

Who this splits for

Choose G​emini 3.8 Live if the product is a conversation. If the value is in how the agent sounds, how it handles being interrupted, how quickly it answers, and how well it works in six languages, it has the best measured profile of anything in its price band and a free tier to prove it on. The 30.1% agentic score only matters if your calls need to end in a completed action rather than a well-handled exchange.

Take MAI Realtime 1 seriously if the product is a transaction — a booking, a claim, a lookup that spans systems — and if you can live with English-only until Microsoft publishes otherwise. At a cent a minute with an unbilled preview, it is the cheapest way to find out whether a vendor-side router beats your application-side one. The risk is not the money. It is that you would be shipping against a model whose only published accuracy figure is its own 98.8%, with no re-run behind it, no τ-Voice score, no language list and no latency distribution — and the only party who can close that gap this month is you, with your own instrumentation.

What neither side has done is publish the number that would settle it: what fraction of real conversations need the second engine. G​oogle's answer is to sell you a second model and let you decide per deployment. Microsoft's is to decide per turn and bill you for it. Both are honest designs. Only one of them is measurable today, and it is not the cheaper one.