Eleven Realtime v1 Max vs SeedRealtime: One Has a Price and Six Benchmarks, the Other Has a Demo

Eleven Realtime v1 Max is a callable speech-to-speech model with a documented endpoint, six published benchmark results and a normalised cost of $4.80 per hour of audio. SeedRealtime is B​yteDance Seed's audio-visual full-duplex model, announced August 5, 2026, and it has no API, no weights, no technical report, no parameter count and no price. The comparison looks decided before it starts — and then you notice that E​levenLabs' model has a benchmark missing too, and that the missing one is the same kind of evidence B​yteDance never published. 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.

That is the real shape of this matchup. Both models are sold on the claim that turn-taking should live inside the model rather than in a voice-activity detector bolted on outside it. Both ask you to trust a behaviour you can only judge by talking to them. E​levenLabs answered that with a third-party evaluation harness and a per-hour cost figure. B​yteDance answered it with one internal preference study. One of those is usable; neither is complete.

What each one actually claims

Eleven Realtime v1 Max merges four stages into one network: recognition, turn-taking, reasoning and voice generation. E​levenLabs' pitch is that tuning them together removes the seams a cascade leaves behind — the dead half-second after a caller stops, the agent that talks across an interruption, the gap between what was heard and what is said back. Max adds the reasoning step to that design and is positioned for complex tasks and multi-step tool calls; its sibling, Eleven Realtime v1 Mini, is the cheap conversational variant, and it is what every new ElevenAgents agent gets by default.

SeedRealtime goes further on modality and less far on everything else. B​yteDance's model unifies audio, video and text in a single end-to-end architecture, which means it can watch, listen and speak at once instead of chaining a vision model to an ASR model to a language model to a voice. The demonstration behaviours are the ones a two-modality system cannot produce: binding a voice to a face in a noisy group conversation so the model tracks who said what, speaking up unprompted when a target object enters frame, museum-guide and device-repair scenarios where the model needs to see what you are looking at.

Audio-plus-video is a harder problem than audio-plus-text, and SeedRealtime is attempting the harder one. That is worth stating plainly before the rest of this piece explains why you still cannot use it.

The evidence, weighed honestly

Artificial Analysis runs both a speech-reasoning test and a conversational-dynamics test on models in this category. It has run them on Eleven Realtime v1 Max. It has not run them on SeedRealtime, and cannot — there is no endpoint to call.

• Speech reasoning (Big Bench Audio) — Eleven Realtime v1 Max 98.8% vs SeedRealtime not published

• Conversational dynamics (Full-Duplex-Bench) — Max 94.1% vs SeedRealtime not published

• Task success rate — Max 87.1% vs SeedRealtime not published

• Arena preference (Elo) — Max 997 vs SeedRealtime not published

• Agentic performance (τ-Voice) — Max not measured vs SeedRealtime not published

• Time to first audio — Max 0.57s vs SeedRealtime no latency figure in milliseconds

• Cost per hour of input audio — Max $4.80 vs SeedRealtime no commercial offer

• Modalities in — Max audio and text vs SeedRealtime audio, video and text

The only row where SeedRealtime leads is the one nobody can buy. The only public number B​yteDance has offered is an end-to-end human evaluation claiming SeedRealtime halves the conversational pacing problems of cascaded ASR-plus-vision-plus-T​TS systems — fewer cut-offs, fewer false triggers, slower and more natural replies. No sample size, no methodology, no annotator count, no latency in milliseconds. A secondary report cites a 12% fluency gain over the team's prior models. That is the entire evidence base.

Now the honest part on the other side. Max's τ-Voice row is blank, which matters because Max is the model E​levenLabs sells for task completion. Its 98.8% reasoning score is measured on 1,000 audio questions, not on a customer-service call with tools and a policy document. So E​levenLabs has six numbers and a hole where B​yteDance has one number and no endpoint — different failure modes, same underlying situation: the thing each model is marketed on is not the thing each model has been measured on.

Why the architecture matters more than the matchup

A year ago, voice agents were a pipeline with an orchestrator on top. Someone had to decide when the user had stopped speaking, route the transcript to a language model, wait for tokens, then hand the text to a synthesiser that had to sound like it had been part of the conversation. Every stage added latency and every handoff added a seam.

Both of these models are bets that the seam is the problem. Eleven Realtime's answer is to merge the audio-side stages — listening, turn-taking, thinking, speaking — into one network and keep the model audio-only. SeedRealtime's answer is to merge those and add sight, so the model can resolve "this one" from a pointing gesture, or notice that the person it is talking to has walked away.

The trade is legible in the specs. Eleven Realtime Max answers in 0.57 seconds and scores 98.8% on a reasoning test with audio-only input. SeedRealtime's claim is broader input and, by implication, more context per turn; its cost is that visual full-duplex is expensive to serve, which is a plausible reason a consumer app is the only place it ships. Neither company has published a latency figure that makes the comparison quantitative, and only one of them has published anything that can be checked by a third party.

The B​yteDance endpoint that exists but is not this model

This is the mistake worth avoiding, because it is easy to make and expensive to discover. Volcano Engine, B​yteDance's cloud arm, documents a separate realtime voice product — the Doubao realtime voice model 3.0 — reachable over a WebSocket endpoint with a console API key, 16 kHz PCM or Opus in and 24 kHz audio out, using a realtime event protocol.

That is a developer API from the same company in the same product area, and it is not SeedRealtime. It is audio-only: no video, no face binding, no visual grounding. It is a distinct model with its own release history. Teams that go looking for B​yteDance's realtime voice API will find that documentation, integrate against it, and believe they are using the model the coverage is about. Volcano Engine has published no date for an endpoint for the audio-visual model. Any listing that offers you SeedRealtime access is offering you something else.

What this means if you have a product to ship

There is no decision to make between these two for production work this quarter. Eleven Realtime v1 Max is callable now, through ElevenAgents or an O​penAI-compatible Realtime API at wss://api.elevenlabs.io/v1/convai/realtime, and SeedRealtime is a research milestone inside a phone app in China.

The decision that does exist is how to build around that. E​levenLabs' Realtime API is WebSocket-only at launch — no WebRTC — and it is a separate product from ElevenAgents, so phone, WhatsApp, knowledge bases, guardrails and monitoring only come with the platform side. Max is also not offered in the India and Singapore data residency environments, and needs a token limit of at least 1,024; Mini, oddly, supports India and Singapore but not Zero Retention Mode or HIPAA, which Max does in the US and EU.

Where OrcaRouter fits is not the voice layer. We do not host eleven_realtime_v1_max, and there is no third-party endpoint for it. What we host is everything an agent of this shape calls around the model: 209 models from 16 upstream providers behind one key, at provider list price with zero markup, with automatic failover when a provider errors or times out. A voice agent's expensive, variable half is the retrieval and the reasoning it delegates; that half is swappable without touching the audio loop, and it is where a week of tuning pays back.

What would change this

An API for SeedRealtime would make this a real comparison, and not only because of the video. A callable endpoint would let Artificial Analysis put the model on the same four-part harness as everyone else, which is the only route to a number that means anything across vendors. A technical report would do most of the same work.

On the E​levenLabs side, the missing piece is a τ-Voice run on Max. When that lands, the field will finally have an apples-to-apples agentic number for the model E​levenLabs markets as the task-oriented one. Until then the summary is unglamorous: one of these models you can call, measure, price and ship, one of these you can only watch, and the gap between those two positions is bigger than the gap between their architectures.