Eleven Realtime v1 Max vs Gemini 3.8 Live: Faster, Sharper on Reasoning, and 5.7× the Price

Eleven Realtime v1 Max starts talking in 0.57 seconds and gets 98.8% of Big Bench Audio right. G​emini 3.8 Live starts talking in 1.18 seconds and gets 92%. On the two axes people quote when they argue about voice models, E​levenLabs wins both — and then loses on almost every axis a finance team cares about, starting with the one where the same benchmark board puts G​emini at $0.84 an hour of audio and Eleven Realtime Max at $4.80. 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.

The two models also disagree about what a voice agent is. G​emini 3.8 Live is a piece of a platform: it has an index score, a documented rate card, 97 languages, SynthID watermarking on every generated second, and two variants — standard and Extended Thinking — that split reasoning from conversation. Eleven Realtime v1 Max is a single model sold on the promise that recognition, turn-taking, thinking and voice were tuned together, and it has no Speech to Speech Index at all because the agentic benchmark behind that index has not been run on it.

That last sentence is the whole argument compressed. One of these models has a complete public scorecard; the other has the best numbers on the two rows it has, and a blank where the hardest row goes.

What the scoreboard actually says

Every figure below comes from Artificial Analysis' Speech to Speech leaderboard, which tests both models on the same harness rather than relaying vendor pages. Where a vendor figure differs, it is flagged.

• Speech reasoning (Big Bench Audio) — Eleven Realtime v1 Max 98.8% vs G​emini 3.8 Live 92%

• Time to first audio — Max 0.57s vs G​emini 1.18s

• Task success rate — Max 87.1% vs G​emini 93.2%

• Agentic performance (τ-Voice) — Max not measured vs G​emini 30.1%

• Conversational dynamics (Full-Duplex-Bench) — Max 94.1% vs G​emini 96.1%

• Arena preference (Elo) — Max 997 vs G​emini 1,078

• Speech to Speech Index — Max not scored vs G​emini 76.0

• Cost per hour of input audio — Max $4.80 vs G​emini $0.84

The pattern is not "E​levenLabs is better at audio, G​oogle is better at everything else." It is narrower and more interesting than that. Max wins the two measurements taken on a fixed audio question set — answer correctly, answer fast — and loses the three that involve a human in a loop or a tool call to complete. Arena preference is people listening and choosing. Task success is whether the right function got called. Conversational dynamics is whether the turn-taking felt right. Those are the production metrics, and G​emini takes all three.

The τ-Voice row deserves a caveat before anyone quotes it. G​emini 3.8 Live's 30.1% is a genuinely poor agentic score — its own Extended Thinking sibling gets 68.6% on the same test, and G​PT-Live-1 (Astra, medium) gets 74.5%. If your agent has to change a booking inside a policy, plain 3.8 Live is not the model you want. But Max's blank is not a zero and should not be read as one. It means the test has not been run.

Where the price gap comes from, and why it may not matter as much as it looks

G​emini 3.8 Live bills at $0.005 per minute of audio in and $0.018 per minute out, with text at $0.75 and $4.50 per million tokens and image or video input at $1.00 per million. Eleven Realtime does not publish a per-model rate at all: in ElevenAgents you pay the standard per-minute agent charge plus tokens at the Realtime model's rate, and through the Realtime API you pay per minute of connected audio plus tokens, with the same token prices on every plan. Artificial Analysis converts both into cost per hour on a fixed question set, which is the only way to compare them honestly, and lands at $0.84 against $4.80.

Two things follow. First, if your volume is measured in thousands of minutes a day, the gap is the decision — 5.7× is not a rounding error, and E​levenLabs has no published lever to close it except the cheaper Mini. Second, the comparison is not static in the way a list price is. G​oogle has cut G​emini audio rates before, and a cut lands on our side the same day when a provider passes list price through, because OrcaRouter marks nothing up. The same is true in reverse: if E​levenLabs publishes a Realtime rate, the number changes without anyone rewriting an integration.

Two different bets about turn-taking

The architectures are not equivalent, and the difference explains the metric split better than any quality claim.

G​emini 3.8 Live is one entry in a family that G​oogle has been shipping for two years.

It processes visual input in near real time, switches languages mid-conversation across 97, and runs tool and API calls in the background so the agent can acknowledge a request and keep talking while the lookup finishes. Its sibling, G​emini 3.8 Live Extended Thinking, reasons and speaks at the same time, narrating progress. Both carry SynthID watermarks and share a model card published as gemini-3-8-audio, and both have documented API IDs: gemini-3.8-live and gemini-3.8-live-extended-thinking.

Eleven Realtime merges the stack rather than coordinating it.

Recognition, turn-taking, the reasoning step and voice generation are one tuned system, which is why it can answer in 0.57 seconds without a separate voice-activity detector deciding when the caller finished. E​levenLabs' own example of the payoff is the backchannel: Mini is documented as detecting whether a speaker has finished or is merely interjecting with an "mm-hmm." That is the behaviour that makes a voice agent stop feeling like a walkie-talkie, and it is the kind of thing a cascaded pipeline struggles with because the VAD has to be told what an interruption is.

The cost of the merged design is control. In ElevenAgents, choosing a Realtime model means E​levenLabs manages the LLM, reasoning effort, thinking budget, parallel tool calls, cascade timeout, T​TS model, expressive mode and ASR provider. You cannot bring a custom LLM, you cannot use Amazon Bedrock, and a workflow node cannot override any of those settings. G​emini gives you the model and lets you keep the pipeline.

The integration question nobody puts in the comparison table

G​emini 3.8 Live is a G​emini API model. Eleven Realtime is sold two ways — inside ElevenAgents, and through a Realtime API that is wire-compatible with O​penAI's event protocol, at wss://api.elevenlabs.io/v1/convai/realtime. If you have an O​penAI Realtime client, changing the endpoint and the model name is most of the migration; the docs include a one-command skill for it.

What you give up for that compatibility is the platform. The Realtime API and ElevenAgents are separate products on different connection protocols, and the API does not carry phone or WhatsApp channels, knowledge bases, guardrails, scripted testing or conversation monitoring. E​levenLabs is also explicit that the Realtime API has no WebRTC at launch — WebSocket only.

That asymmetry is where a routing layer fits, and it fits on the G​emini side better than on the E​levenLabs side. OrcaRouter does not host eleven_realtime_v1_max; there is no third-party endpoint for it, and we will not pretend otherwise. What OrcaRouter does host is the rest of a G​emini-shaped stack: google/gemini-3.8-flash and the other G​emini models behind one key, 209 models from 16 providers total, at provider list price with zero markup, with automatic failover when a provider errors. If you build the voice layer on G​emini, the retrieval and post-call analysis around it can be swapped and A/B-tested on your own traffic rather than committed to from a leaderboard.

Which one to pick

Pick G​emini 3.8 Live if the agent has to finish a transaction, if you need the price under a dollar an hour, if you need visual input, or if you need 97 languages on one line. It has the complete public scorecard, the documented rate card and the cheaper minute. Its agentic weakness is real but measurable, and its Extended Thinking sibling is the fix if task completion is the bottleneck.

Pick Eleven Realtime v1 Max if the conversation itself is the product — a tutor, a character, a support line where being cut off is the complaint — and if your workload is reasoning-heavy audio rather than long connected calls. You are buying the fastest strong reasoner on the board and the best-measured turn-taking, at a price premium you should model before you commit.

Do not pick on the index score alone. Max does not have one, and that is not a knock on Max's quality; it is a statement about how much of Max's behaviour is still unmeasured. The one number that would change this comparison is a τ-Voice run on Max. If it lands above G​emini's 68.6% — the Extended Thinking variant, not the standard one — E​levenLabs' premium starts to look like an insurance policy rather than a tax. Until it lands, the faster model is the one with more to prove.