--- title: OpenAI Details GPT-Live’s Architecture for Continuous... description: OpenAI published an engineering deep dive into GPT-Live, its architecture for continuous, stateful voice interaction. The design separates a latency-critical... canonical: https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction | daily.dev og:description: OpenAI published an engineering deep dive into GPT-Live, its architecture for continuous, stateful voice interaction. The design separates a latency-critical... og:url: https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop og:image: https://api.daily.dev/og/posts/um8nuVvOp.png og:image:alt: OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction **[InfoQ](https://daily.dev/sources/infoq)** · 5 min read · 1 upvotes · 0 comments ## Summary OpenAI published an engineering deep dive into GPT-Live, its architecture for continuous, stateful voice interaction. The design separates a latency-critical live path (media pipeline and inference loop) from application logic (delegation, tool use, persistence) via an asynchronous RPC boundary. Sessions use dedicated stateful inference that can migrate context between model instances when capacity drains or context limits are reached. The system retains WebRTC as its media foundation, adding WARP (WebRTC Abridged Roundtrip Protocol) improvements like SPED, DTLS 1.3, and SNAP, plus Instant Connect, to cut startup latency without replacing the transport layer. Before launch, OpenAI ran a 'silent test' that fed real voice traffic through the system while discarding output, revealing load-related issues—such as GPU/CPU colocation problems in certain regions—that synthetic load tests missed. Justin Uberti, Head of Realtime AI at OpenAI, discussed these trade-offs in an interview. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why did OpenAI keep WebRTC instead of switching to RTP over QUIC for GPT-Live's voice architecture? WebRTC was kept because it provides a battle-tested low-latency media stack with built-in error recovery, while RTP over QUIC currently only covers the transport layer and lacks features like GCC congestion control and RTT-aware path selection. OpenAI instead extended WebRTC with WARP improvements (SPED, DTLS 1.3, SNAP) and Instant Connect, which existing WebRTC apps benefit from without code changes. _Teams weighing WebRTC against newer transport stacks can follow real-world architecture trade-offs like this on daily.dev._ ### How does GPT-Live handle session state when a voice conversation needs to move to a different model instance? GPT-Live uses dedicated, stateful inference where each session reserves capacity on its assigned instance, but session context can migrate in real time to another instance when the original instance is draining capacity or the conversation nears its context limit. This lets OpenAI steer new sessions to available capacity while scaling instances up and down based on demand. _Engineers designing stateful real-time systems can track architecture writeups like this on daily.dev._ ### What did OpenAI's silent test reveal about GPT-Live that synthetic load testing missed? The silent test, which fed authentic voice traffic through GPT-Live in read-only mode while discarding output, revealed that performance degraded under load in ways synthetic tests hadn't predicted. Specifically, in certain regions some GPUs were not colocated with the CPUs feeding them, introducing unexpected latency that was fixed before launch. _Anyone validating real-time systems against production traffic can find case studies like this on daily.dev._ --- Tags: [#career](https://daily.dev/tags/career), [#backend](https://daily.dev/tags/backend), [#openai](https://daily.dev/tags/openai), [#webrtc](https://daily.dev/tags/webrtc) [View this post on daily.dev](https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction","url":"https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop"},"datePublished":"2026-09-02T12:24:24.034Z","dateModified":"2026-09-02T12:29:55.388Z","description":"OpenAI published an engineering deep dive into GPT-Live, its architecture for continuous, stateful voice interaction. The design separates a latency-critical...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/9d1afe1b7bc1f597a6b9371b67ce2367?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/9d1afe1b7bc1f597a6b9371b67ce2367?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"InfoQ","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"InfoQ","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/afc3bced3e1e4b188dd9127017a60e0c","url":"https://daily.dev/sources/infoq"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"career,backend,openai,webrtc","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"InfoQ","item":"https://daily.dev/sources/infoq"},{"@type":"ListItem","position":3,"name":"OpenAI Details GPT-Live’s Architecture for Continuous Stateful Voice Interaction"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/openai-details-gpt-live-s-architecture-for-continuous-stateful-voice-interaction-um8nuvvop#faq","mainEntity":[{"@type":"Question","name":"Why did OpenAI keep WebRTC instead of switching to RTP over QUIC for GPT-Live's voice architecture?","acceptedAnswer":{"@type":"Answer","text":"WebRTC was kept because it provides a battle-tested low-latency media stack with built-in error recovery, while RTP over QUIC currently only covers the transport layer and lacks features like GCC congestion control and RTT-aware path selection. OpenAI instead extended WebRTC with WARP improvements (SPED, DTLS 1.3, SNAP) and Instant Connect, which existing WebRTC apps benefit from without code changes. Teams weighing WebRTC against newer transport stacks can follow real-world architecture trade-offs like this on daily.dev."}},{"@type":"Question","name":"How does GPT-Live handle session state when a voice conversation needs to move to a different model instance?","acceptedAnswer":{"@type":"Answer","text":"GPT-Live uses dedicated, stateful inference where each session reserves capacity on its assigned instance, but session context can migrate in real time to another instance when the original instance is draining capacity or the conversation nears its context limit. This lets OpenAI steer new sessions to available capacity while scaling instances up and down based on demand. Engineers designing stateful real-time systems can track architecture writeups like this on daily.dev."}},{"@type":"Question","name":"What did OpenAI's silent test reveal about GPT-Live that synthetic load testing missed?","acceptedAnswer":{"@type":"Answer","text":"The silent test, which fed authentic voice traffic through GPT-Live in read-only mode while discarding output, revealed that performance degraded under load in ways synthetic tests hadn't predicted. Specifically, in certain regions some GPUs were not colocated with the CPUs feeding them, introducing unexpected latency that was fixed before launch. Anyone validating real-time systems against production traffic can find case studies like this on daily.dev."}}]} ```