GPT-Live-1 vs SeedRealtime: SeedRealtime Is the More Ambitious Model, and You Cannot Buy It

Comparing GPT-Live-1 and SeedRealtime on capability produces an uncomfortable answer, because the two models are not competing on the same terms. GPT-Live-1 is Ope​nAI's full-duplex voice model, shipped inside ChatGPT on July 8, 2026 and opened to developers through the Live Sessions API on September 10, 2026, with 12 voices, a published per-minute rate, and one significant hole: image and video input are explicitly unsupported. SeedRealtime, released by ByteDance's Seed team on August 5, 2026, is built entirely around that hole. It is a native audio-visual full-duplex model that watches, listens and speaks in a single end-to-end architecture — and it is available only inside the Doubao consumer app, with no public API, no open weights, no technical report and no published parameter count.

What each one can actually perceive

The cleanest way to see the gap is to ask what each model knows about the room it is in.

GPT-Live-1 hears. That is the whole input surface. Audio and text go in, audio and text come out, and Ope​nAI's documentation lists image and video as not supported. It handles the conversational mechanics of hearing extremely well — it decides many times a second whether to keep listening, pause, interrupt or call a tool, and it maintains full duplex so it keeps processing incoming audio while it is speaking. It also returns speech-recognition transcripts alongside the audio, which is the sort of unglamorous detail that saves a week of integration work.

SeedRealtime hears and sees, in one model rather than a pipeline. ByteDance describes a unified architecture that handles audio, video and text together, resolving ambiguity with visual context — distinguishing homophones, understanding what "this" refers to from scene, gesture, gaze and prior action — and running perception, understanding, decision-making and expression in parallel rather than in sequence. ByteDance's published demonstrations include a noisy group dinner where the model has to bind voices to identities, a museum visit, correcting an espresso workflow, and suppressing interference in an airport while holding off-screen memory and doing an online lookup.

• Inputs — GPT-Live-1 audio and text only, image and video explicitly unsupported vs SeedRealtime audio, video and text fused in one model

• Turn-taking — GPT-Live-1 decides internally, many times per second vs SeedRealtime moves turn-taking inside the model and removes the external voice-activity detector entirely

• Proactive behaviour — GPT-Live-1 responds, delegates and calls tools vs SeedRealtime is described as speaking up unprompted when a target object appears in view

• Availability — GPT-Live-1 generally available in Ope​nAI's API at $0.05 a minute vs SeedRealtime free inside Doubao, with no API and no timeline for one

The evidence quality runs the opposite way to the ambition

This is where the comparison stops flattering ByteDance.

Ope​nAI publishes numbers. They are its own, they compare GPT-Live-1 against GPT-Realtime-2.1 rather than against a competitor, and no independent lab has reproduced them — 80.1% on full-duplex interactivity against 45.4%, turn-taking latency cut from 1.4 seconds to 0.8, tool-calling accuracy from 60% to 87%. Vendor-run and unreproduced is a real caveat, and it is a caveat you can at least attach to a figure.

ByteDance publishes almost nothing. The central claim about SeedRealtime is an end-to-end human evaluation saying it halves audio-visual 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 methodology, no exact figure for the improvement, and no latency number at all. There is also no parameter count and no technical report.

The result is that the model with the more interesting architecture is the one you can evaluate least. That is not the same as saying it performs worse — the demonstrations are genuinely striking, and the engineering write-up around chunked audio-visual input and streaming generation suggests a real system rather than a demo — but it does mean that any comparison between these two on quality is currently a comparison between a documented model and an impressive video.

Why the API gap is not a detail

SeedRealtime has been fully rolled out inside Doubao since August 5, 2026, in both voice and video, free of charge. It has no Volcano Engine or BytePlus endpoint, no paid tier, no published quota, and no stated timeline for opening developer access. ByteDance's roadmap mentions lower latency, sharper speaker tracking and tool-connected tasks; it does not mention a date.

There is an access caveat that compounds it. Doubao is a mainland-China-first product and typically requires a mainland mobile number to register, with no localised English version announced. For a team outside China, SeedRealtime is not merely unlaunched as an API — it is not reachable as a product either.

Set that against GPT-Live-1's own integration burden and the trade becomes concrete. Ope​nAI's API is narrow in ways that surprise people: one model ID, one endpoint at v1/live/sessions, WebRTC transport with event handling on a data channel, and no path through Chat Completions, Responses, Realtime, Batch or the fine-tuning endpoints. Rate limits are denominated in concurrent sessions — 25 on Tier 1, rising through 50, 200, 300 and 500 — rather than requests per minute, and the Free tier cannot call the model at all. That is a new integration, and it is still an integration you can start this afternoon.

Two answers to the same delegation problem

Both models reject the old cascaded design, where speech recognition, a language model and speech synthesis run in sequence with roughly 1.7 seconds of dead air between turns. They reject it differently, and the difference tells you which one is built for a phone call and which is built for a room.

GPT-Live-1 splits the work vertically. A fast voice layer holds the conversation; anything heavier — web search, deeper reasoning, agentic work — is handed to a separate reasoning model through the Responses API while the conversation continues. Ope​nAI's published WebRTC example wires that backend to GPT-5.6 Terra. The consequence is that the thinking half is an ordinary text-model call, priced per token, and therefore something you can choose, swap and route independently of the voice layer. For a support line, that is the right shape: the voice is the interface and the reasoning is the product.

SeedRealtime keeps everything in one network, which is what lets it look at a gesture while it is mid-sentence. It is also what makes it impossible to decompose — you cannot swap the reasoning half, because there is no reasoning half, and you cannot run it anywhere, because there is nowhere to run it.

If you are building a voice agent today, that distinction is the whole decision. Only one of these two is a component you can put in an architecture. GPT-5.6 Terra, the backend in Ope​nAI's delegation example, is served through OrcaRouter at $2.00 per million input tokens and $12.00 per million output with a 1M-token context and both /v1/chat/completions and /v1/responses exposed — alongside roughly 190 other models on one key, so the reasoning tier behind a voice agent can be split, priced and failed over without touching the audio path. Neither GPT-Live-1 nor SeedRealtime is served by OrcaRouter; they are single-source from their own vendors, and no router can change that.

What would change the answer

Three things, in order of likelihood.

• A ByteDance developer API. If SeedRealtime shows up on Volcano Engine or BytePlus with a published rate, it stops being a case study and becomes a genuinely differentiated product — audio-visual full duplex with no hosted competitor in the West. That is the signal to watch, and it has not appeared.

• Vision arriving in a hosted voice API. GPT-Live-1's documentation is explicit that image and video are unsupported, which reads less like an oversight than like a deliberate first-release boundary. If Ope​nAI ships the video surface it has been demonstrating elsewhere in ChatGPT, the capability gap that makes SeedRealtime interesting narrows considerably.

• Any independent measurement of SeedRealtime. A third-party latency figure or a parameter count would let the two be compared on something other than ambition.

The practical verdict for anyone building now is short. GPT-Live-1 is the only one of these two you can deploy, and at $0.05 a minute billed by the second, with twelve voices and a documented endpoint, it is a reasonable default for conversational voice. SeedRealtime is the more interesting model and may well be the more capable one; it is also, for the moment, a demonstration of what a converged audio-visual architecture looks like rather than a product you can put in front of a customer. File it under what to watch, not what to build on.