Generated title card titled 'GPT-Live-1.1 vs SeedRealtime' with the subtitle 'One sells minutes, one sells nothing', showing a left card labelled GPT-Live-1.1 with $0.05 per active minute, $0.02 per standby minute and a public API, and a right card labelled SeedRealtime with no public API, no open weights and availability inside Doubao only.
Guides & Insights

GPT-Live-1.1 vs SeedRealtime: One Sells Minutes, One Sells Nothing

Author

jinhao.song

Date Published

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

Put GPT-Live-1.1 and SeedRealtime side by side and the comparison immediately stops being about quality. GPT-Live-1.1 is a full-duplex voice model you can call today, rent by the second, and put in a product; it appeared on the O​penAI API on November 19, 2026, five weeks after it replaced GPT-Live-1 inside ChatGPT Voice. SeedRealtime, which ByteDance's Seed team announced on August 5, 2026, is the more interesting architecture on paper — a native audio-visual full-duplex model that watches and listens and speaks in one system — and it is not something you can call. No public API, no open weights on Hugging Face, no parameter count, no technical report with numbers in it.

So the honest framing is not "which is better." It is: one of these is a product with a rate card, and the other is a demonstration of a capability the first one has explicitly declined to ship. Both statements are true at once, and which one matters depends entirely on whether you are building something.

The input surface is not the same shape

GPT-Live-1.1 hears. That is the whole story of its input. The model page lists audio and text as input modalities and audio and text as output, and nothing else — no image, no video, no document. What it does with hearing is genuinely unusual: it is full duplex, meaning it listens while it speaks rather than alternating turns, and it decides many times a second whether to keep listening, pause, interrupt or hand work to a backend. The version bump's only stated capability change is in this layer: compared with GPT-Live-1 it "transcribes speech more accurately and maintains a more consistent accent and voice across languages." No number is attached to that claim.

SeedRealtime hears and sees, in one model rather than a pipeline. ByteDance's launch post describes a unified architecture that fuses audio, video and text, running perception, understanding, decision-making and expression in parallel over continuous streams. The concrete consequences it lists are the ones a text-only voice model structurally cannot reach: resolving a homophone by looking at the scene, working out what "this" refers to from gesture and gaze, and speaking up unprompted when a target object appears in frame. The demonstrations are specific — binding names to faces across a noisy dinner table, translating a Chinese-only menu for a foreign diner, noticing whole beans going into a portafilter and saying so before extraction, holding a museum exhibit in view until the user's requested piece arrives and then reminding them.

• Inputs — GPT-Live-1.1 audio and text only vs SeedRealtime audio, video and text fused in a single model

• Turn-taking — GPT-Live-1.1 decides internally at high frequency vs SeedRealtime moves turn-taking inside the model and removes the external voice-activity detector

• Proactive behaviour — GPT-Live-1.1 responds, delegates and calls tools vs SeedRealtime is described as speaking up unprompted when something appears in view

• Where it runs — GPT-Live-1.1 on v1/live/sessions at a published per-minute rate vs SeedRealtime inside ByteDance's Doubao app

• Evidence — GPT-Live-1.1 has third-party scores for its predecessor on Artificial Analysis vs SeedRealtime has an end-to-end human evaluation with no sample size and no figure

Generated two-column scoreboard titled 'GPT-Live-1.1 vs SeedRealtime — the scoreboard' with six matched rows: inputs (audio and text vs audio, video and text fused), turn-taking (decided inside the model vs moved into the model), proactive behaviour, where each runs (v1/live/sessions vs inside Doubao only), evidence (index scores for GPT-Live-1 only vs a vendor human evaluation with no figure), and price ($0.05/$0.02 per minute vs no developer product).

What you actually pay, and to whom

GPT-Live-1.1 bills voice sessions by duration, per second, at $0.05 per active minute. Since November 19, 2026, a WebRTC session can also be configured to pause inference during inactivity and bill at $0.02 per minute while it waits, resuming when someone speaks. Backend model and tool usage — the reasoning and tool work the voice layer delegates — is charged separately on the token meter of whichever model you name. Creating a WebRTC session bills 15 seconds of voice duration at initialisation, credited back against duration once the session runs.

SeedRealtime's price to a developer is not a number, because there is no developer product. It is available inside Doubao, the consumer assistant, where ByteDance says it has been "fully rolled out" since the August launch. That is a real deployment at real scale, and it is not an API. Whether ByteDance's model-as-a-service platform will carry it, and at what rate, is not stated in the launch post or on the Seed site as read on November 20, 2026.

The asymmetry has a cost consequence beyond the obvious one. A model you rent by the second can be budgeted, compared and switched off. A model inside someone else's consumer app cannot be any of those things, and the comparison between the two therefore has no exchange rate. You are not choosing between $0.05 a minute and something cheaper; you are choosing between a line item and a demo.

Where the evidence thins out

GPT-Live-1.1 itself has no independent score. Artificial Analysis' speech-to-speech index, read on November 20, 2026, lists GPT-Live-1 twice — 83.2 with an Astra-class backend at medium effort and 80.1 with a Sol-class backend at low — and does not yet carry 1.1. That is a day-old API model, so the gap is expected rather than damning, but it does mean the accent-consistency improvement is, this week, a vendor sentence. What the index does show is that configuration moves the number as much as the version does: the same model scores 83.2 or 80.1 depending on the backend you attach. Any review quoting one GPT-Live figure without naming its backend has told you very little.

SeedRealtime's evidence is thinner in a different way. ByteDance publishes one headline result — an end-to-end human evaluation in which the model "reduces audio-visual conversational pacing issues by half" compared with cascaded ASR-plus-vision-plus-TTS systems, with fewer mid-sentence cut-offs, fewer sluggish replies after a pause, and fewer false triggers from background chatter. There is no sample size, no methodology, no confidence interval and no latency figure. There is also no parameter count and no technical report. That is not evidence of a weak model — the demonstrations are specific enough to suggest a real system — but it is a claim you cannot check, and it does not appear on any independent leaderboard.

Screenshot of the ByteDance Seed blog post announcing SeedRealtime, showing the 2026-08-05 publication date, the Audio category label, and the three stated breakthroughs of the model.

Note what that means for the comparison specifically: the two models are not on a common scale. One has third-party numbers for the previous version and a published rate; the other has a vendor sentence and no way to buy it. Ranking them against each other would be inventing a measurement.

Why the API gap is the whole argument

It is tempting to treat "no API" as a temporary state that will resolve itself and therefore as a detail. For a build decision, it is the opposite: it is the only fact that fully determines the answer.

If you are shipping a voice product in the next quarter, SeedRealtime is not on your list of options, and no amount of architectural ambition changes that. GPT-Live-1.1 is on the list, with a documented session configuration, a published rate, a standby mode that makes idle connections cheap, and a delegation model that lets you keep the reasoning in a backend you control. The one thing it will not do is look at your user's camera.

That last limitation is worth pricing honestly, because it is where a team's requirements actually split. A support line, an in-car assistant, a language tutor, a scheduling agent — none of these need video, and all of them benefit from a model that handles interruption and delegation well at a known cost. An agent that watches a repair, reads a menu, or follows a demonstration needs the modality SeedRealtime has and GPT-Live-1.1 does not, and for that workload the answer today is to wait, or to build the cascade ByteDance is arguing against.

Generated decision card titled 'Which voice workload fits which model', with a left column headed 'GPT-Live-1.1 can take this today' listing support lines, in-car assistants, language tutors and scheduling agents, and a right column headed 'This needs a camera' listing repair guidance, menu reading and demonstration following, captioned 'SeedRealtime shows these; no public API yet.'

Where a router fits is the delegation half. OrcaRouter does not serve GPT-Live-1.1 or SeedRealtime — neither is a chat model, and the catalogue carries no full-duplex voice endpoint. It does serve the backend the voice layer hands work to: gpt-6-luna, gpt-6-sol and the rest of the GPT-6 family are available behind one key at provider list price with no markup, with automatic failover across providers and a routing DSL for composing several models into one call. If your Live session delegates to a backend, that backend is the part you can route today.

What to watch, and what to do now

Two things would change this comparison, and neither has happened. A public SeedRealtime endpoint — on ByteDance's own platform or anywhere else — would turn a demonstration into an option overnight, and would immediately raise the question of whether visual context is worth the integration work. An independent measurement of GPT-Live-1.1 on the speech-to-speech index would tell you whether the multilingual accent claim is real, and whether the 83.2 that GPT-Live-1 posts with an Astra backend carries over.

Until then, the decision is unusually clean. If you need to ship voice, GPT-Live-1.1 is callable, priced, and — with standby — cheaper to leave connected than it was a week ago. If you need a model that sees, no amount of capability on ByteDance's side converts into an API key, and the responsible thing is to say so rather than to rank a product against a preview.

Compared in this article1

Detected from this article · Benchmarks: Artificial Analysis · updated daily