Grok Voice 3.0 vs SeedRealtime: One Is the Best Voice Model on the Board, the Other Is Not on the Board

This comparison has an unusual shape. Grok Voice 3.0 is the highest-scoring speech-to-speech model Artificial Analysis measures, at 86.9 on the Speech to Speech Index as of 13 October 2026, priced at $0.08 per minute and reachable today through SpaceXAI's realtime API. SeedRealtime — formally "SeedRealtime: An Audio-Visual Full-Duplex LLM", released by ByteDance's Seed team on 5 August 2026 — does not appear on that board in any form. No index score, no reasoning component, no conversational-dynamics figure, no arena Elo, no task success rate, no time to first audio. Not a low score. A blank row, because the model has never been entered.

That blank is the honest starting point, because it is not an oversight by the board. SeedRealtime has no public API, no model ID, no open weights, no technical report, no parameter count and no published price. It runs inside ByteDance's Doubao consumer app, free, in voice and video, and has done since release. There is nothing to measure because there is no endpoint to measure, and there is no way to call it because there is nothing to call.

Where they genuinely differ, and it is not quality

Set the scoreboards aside and the two models are doing different jobs in different rooms.

Grok Voice 3.0 is an audio model. It hears and speaks over a duplex WebSocket, decides many times a second whether to keep listening, pause, interrupt or invoke a tool from the tools array in its session update, and it holds five measured component scores plus the top arena position. Its predecessor generation was marketed on word error rate across 24 languages and 1.5 to 2.0 times lower error than dedicated transcription models — vendor-reported, never independently reproduced, and not restated in the newer launch material.

SeedRealtime hears and sees in one network. ByteDance describes a unified framework in which perception, understanding, decision-making and response generation happen inside a single model rather than in a cascade, with turn-taking decided internally and no external voice-activity detector at all. Its published demonstrations include a noisy group dinner where the model has to bind voices to identities, a museum visit where it volunteers information when a target exhibit comes into view, an espresso workflow it corrects in real time, and an airport scenario where it ignores unrelated chatter, recalls a departure board from earlier in the session and goes online for a baggage carousel number.

• Inputs — Grok Voice 3.0 audio and text, with tools; SeedRealtime audio, video and text fused in one model.

• Turn-taking — Grok Voice 3.0 uses server-side voice activity detection configured per session; SeedRealtime removes the external VAD and moves the decision inside the model.

• Proactive behaviour — Grok Voice 3.0 responds and calls tools; SeedRealtime is described as speaking up unprompted when a relevant object appears in view.

• Latency — Grok Voice 3.0 publishes 0.81 seconds to first audio; SeedRealtime publishes no latency figure at all, only a roadmap goal to compress end-to-end latency.

• Price — Grok Voice 3.0 $0.08 per minute plus $0.004 per text input; SeedRealtime has no price and no commercial offer, and is free inside a consumer app.

• Reachability — Grok Voice 3.0 is live for developers with a documented endpoint and an ephemeral-token path for client-side connections; SeedRealtime is mainland-China-first through Doubao, which typically requires a mainland mobile number to register, with no localised English product announced.

• Evidence — Grok Voice 3.0 has five independent component measurements plus a composite; SeedRealtime has one internal human evaluation.

The evidence problem, stated precisely

The single quantitative claim ByteDance makes about SeedRealtime is that end-to-end human evaluations show it halves conversational pacing problems relative to cascaded ASR-plus-vision-plus-TTS systems — fewer cut-offs mid-sentence, fewer slow replies after a pause, fewer false triggers from background chatter. There is no sample size, no annotator count, no benchmark name, no methodology and no exact figure. The claim is that the problem got roughly half as bad, and it comes from the team that built the model.

Compare that with what is knowable about Grok Voice 3.0. Its 86.9 composite decomposes into a reasoning figure, a conversational-dynamics figure, an agentic score, an arena preference Elo and a task success rate, each measured under a published harness, with the underlying board readable by anyone. That does not make the numbers beyond dispute — a composite index is a weighting choice, and the arena component is frozen at the point a model becomes eligible for publication rather than tracking live Elo — but it means an outside team can check the arithmetic and disagree with the conclusion rather than being unable to form one.

Both models also carry vendor claims about their own production deployments, and both sets are unverifiable in the same way. SpaceXAI reports Starlink call centres handling 20,000 customer calls a day in multiple languages with roughly 70% resolved without a human, and millions of daily conversations across Tesla and SpaceX. ByteDance reports full rollout in Doubao across voice and video. These are the strongest available signals that the systems work in the real world, and simultaneously the weakest available evidence for a reader, since neither is auditable from outside and both describe services rather than models.

What you can build on, and what you cannot

Here is the part that resolves the comparison regardless of capability. Only one of these two is a component. The other is a demonstration of what a converged architecture looks like.

Grok Voice 3.0 slots into a stack with a model string, or with nothing at all if you deployed against grok-voice-latest, which was moved onto it automatically on 14 October. The same realtime endpoint, the same session parameters, the same event names, the same built-in voice set and custom-voice endpoint as the generation it replaces. Session-scoped tools let the agent call your backend mid-conversation. There is an integration cost, and it is measured in minutes.

SeedRealtime cannot be decomposed, and that is architectural rather than commercial. There is no reasoning half to swap, because the thinking is inside the same network as the listening and the looking; there is no hosted surface to route to, no quota to buy, no key to obtain. Its capability is real and its demonstrations are the most striking in the category. It is also, for anyone outside ByteDance's consumer market, not a product you can put in front of a customer this quarter.

One warning worth stating plainly, because it is a live failure mode: at least one third-party blog currently describes SeedRealtime as available through a model router. There is no endpoint to route to, so anything advertising SeedRealtime access is advertising something else. Relatedly, searches for ByteDance's realtime voice API surface Volcano Engine's end-to-end realtime speech model, which is a different, audio-only system — if you have been comparing that model's published quota and pricing against Grok Voice 3.0, you have been comparing the wrong product.

Where the router actually fits

Neither model is served by OrcaRouter, and it is worth being blunt about why the framing usually gets this wrong. Grok Voice 3.0 is single-source from xAI and SeedRealtime is not sold by anyone. No routing layer changes either fact, and a platform claiming otherwise is describing a capability it does not have.

What a routing layer changes is the text tier underneath the voice. Every one of these deployments is a voice model plus a language-model stack: the tool backend the voice model calls, the classifier that reads the transcript afterward, the summariser that writes the record, the retrieval that answers what the voice layer had to hand off. That work is request-metered and it is where a meaningful slice of per-call cost accumulates. It is what we run — 200-plus models behind a single key at provider list price with zero markup, automatic failover across providers, and a routing DSL for composing several models into one call — and it is the half of a voice product that you can hold constant while the audio model underneath it changes, which given how casually the audio layer gets swapped is a property worth having.

The verdict

If you are building a voice agent now, Grok Voice 3.0 is the pick, and the reason is not the index. It is that it holds the highest composite and the top arena position, it publishes the fastest start-to-speak time in the frontier group at 0.81 seconds, it costs the same per minute as the model it retires, and it is live with a documented endpoint today. That is four separate reasons, and none of them require trusting a number that nobody can check.

Choose SeedRealtime — or rather, wait for it — only if your product depends on the specific thing it does that nothing else does: reading a room. Audio-visual full duplex with eyes on the scene is a capability no hosted Western voice model currently offers, and ByteDance's demonstrations of it are convincing. The trigger to watch is a ByteDance developer API on Volcano Engine or BytePlus with a published rate, which would turn a case study into a product with no hosted competitor. Until that appears, the model stays on the watch list rather than in the architecture, and any article claiming to compare it to a leaderboard model is comparing a scoreboard to a video.

Comparisons like this one are also where the failure modes show up: 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.