[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gpt-live-1-api-launch":3,"topics-all":38,"news-related-0f466258-290e-4c6b-b471-2169ba6a393f":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"0f466258-290e-4c6b-b471-2169ba6a393f","GPT-Live-1 进 API:全双工语音层 0.05 美元一分钟,推理外包给 GPT-6 Astra","9 月 10 日 OpenAI 把全双工语音模型 GPT-Live-1 开放进 API:一个模型同时听说、原生处理打断,深层推理可委托给 GPT-6 Astra 等后端;语音层每分钟 0.05 美元,支持电话场景。Speak 早期评测称思考停顿期误打断减少近 80%。","7 月 8 日,OpenAI 在 ChatGPT 里首发了全双工语音模型 GPT-Live-1;两个月后的 9 月 10 日,它进了 API,第三方开发者第一次可以把这套「能同时听和说」的语音层装进自己的应用。这不是一次简单的接口开放——OpenAI 同时给出了定价、电话场景支持和一整套「语音层与推理层分离」的架构主张。\n\n## 一个模型干掉三段式管线\n\n传统语音智能体是三段式拼接:语音转文本(STT)、大模型推理、文本转语音(TTS)。每一段交接都在加延迟,也都在制造事故点——用户什么时候打断、停顿、改口,全要开发者手写规则去兜。GPT-Live-1 的做法是把听和说收进同一个模型,由它直接对输入和输出音频联合推理,打断、应答、背景噪音都在模型内部处理。官方公告的说法是,这避免了级联架构的延迟与脆弱交接。模型还原生输出 ASR 转写文本、支持关键词偏置和回合检测——虽然它本身不是回合制模型,但给习惯回合制管道的开发者留了兼容层。落地效果有两个数字:语言学习应用 Speak 在早期评测中发现,思考停顿期的误打断减少了近 80%;一位医疗健康公司的联合创始人兼 CTO Tony Stoyanov 称,换成 GPT-Live-1 后代码库简化了 80%,删掉了 2.3 万行代码。\n\n## 推理外包:语音层按分钟卖,大脑自选\n\n更有意思的是架构分工。GPT-Live-1 自己只负责「对话」,深层推理和工具调用可以委托给后端文本模型——官方例子是 GPT-6 Astra,也可以是第三方模型。语音层定价每分钟 0.05 美元,后端模型和工具费用另算。对开发者来说,这意味着可以按任务混搭:官方给的场景是,排期、订单更新这类高频任务配轻量模型,复杂客诉配重推理模型。OpenAI 还补齐了电话(telephony)支持——全双工语音智能体可以直接部署到电话场景;长会话的上下文保持、通过系统提示词控制语气语速、覆盖更多口音和语言的新音色,也都列进了发布清单。公告页面上亮出名字的客户是 Yelp、Speak、Fin 和 Cognition,Cognition 的联合创始人 Walden Yan 说,配上 Devin,和 AI 工程师协作开始像和同事对话。\n\n## 数字与冷思考\n\n跑分方面,OpenAI 称 GPT-Live-1 在 Full Duplex Bench 上比 GPT-Realtime-2.1 高出 30 个百分点,回合接管延迟和交互行为提升明显;配上 GPT-6 Astra(中等推理力度)在衡量语音智能体端到端任务能力的 Tau3 榜单排名第一。独立媒体的报道也在转述这组官方数字,但目前都出自 OpenAI 自己的评测口径。值得留意的还有计费结构:按分钟计费加后端另算,做长时电话客服的团队需要把成本模型算清楚。\n\n所以呢?语音交互的护城河正在从「谁的声音自然」转向「谁的架构简单」。当听、说、打断在一个模型里闭环,而推理可以插拔,语音智能体的竞争点就回到了后端智能与工作流集成——那恰恰是 OpenAI 顺便一起卖的东西。原文公告:https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-gpt-live-1-in-the-api\u002F\n","https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-gpt-live-1-in-the-api\u002F","15975962-b5fe-49e5-ae68-687ba6cb7015",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"e676a5cf-1f24-472f-a765-86fa21a1bc3c","ai-model",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"baf131c1-687a-49f4-87f6-4dd87c1c692f","gpt",{"id":19,"name":20,"slug":20,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":22,"name":23,"slug":23,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"80083c69-f5d5-4e75-b980-f4899a61a52d","en","GPT-Live-1 lands in the API: full-duplex voice at $0.05\u002Fmin","GPT-Live-1, OpenAI's full-duplex voice model, lands in the API: listens and speaks in one model, delegates reasoning, $0.05\u002Fmin.","On July 8, OpenAI first shipped GPT-Live-1 inside ChatGPT as its full-duplex voice model. Two months later, on September 10, it landed in the API — and for the first time, third-party developers can bolt that \"listens while it speaks\" voice layer into their own apps. This is more than an endpoint opening: OpenAI also published pricing, telephony support, and a clear architectural bet on separating the voice layer from the reasoning layer.\n\n## One model replaces the three-stage pipeline\n\nA traditional voice agent is a cascade: speech-to-text, an LLM, then text-to-speech. Every handoff adds latency and creates failure points — developers end up hand-coding rules for interruptions, pauses, and mid-sentence changes of mind. GPT-Live-1 collapses listening and speaking into a single model that reasons over incoming and outgoing audio together; the official announcement says this avoids the latency and brittle handoffs of chained STT–LLM–TTS architectures. It natively emits ASR transcripts and response text, supports keyword biasing, and — although it is not a turn-based model — offers native turn detection for teams still building around explicit turns. Two numbers stand out from early deployments: language-learning app Speak found interruptions during thinking pauses dropped by almost 80% versus previous turn-based systems, and Tony Stoyanov, co-founder and CTO at a healthcare company, says switching to GPT-Live-1 simplified his codebase by 80% and removed 23,000 lines of code.\n\n## Outsourced reasoning: voice by the minute, brain of your choice\n\nThe more interesting move is the division of labor. GPT-Live-1 handles the conversation itself, while deeper reasoning and tool calls can be delegated to a backend text model — the official example is GPT-6 Astra, but third-party models work too. The voice layer is priced at $0.05 per minute, with backend models and tools billed separately. The scenario OpenAI sketches: pair a lightweight model for high-volume tasks like scheduling or order updates, and a heavier reasoning model for complex customer issues. Telephony support rounds out the release — full-duplex voice agents can be deployed for phone calls — along with long-session context retention, tone and pace control through the system prompt, and new voices across more accents, dialects, and languages. The customers named on the announcement page are Yelp, Speak, Fin, and Cognition; Cognition's co-founder Walden Yan said that with Devin, working with an AI engineer starts to feel like collaborating with a teammate.\n\n## The numbers, and a cooler head\n\nOn benchmarks, OpenAI says GPT-Live-1 improves Full Duplex Bench performance by 30 percentage points over GPT-Realtime-2.1, with large gains in turn-taking latency and interactive behavior, and that paired with GPT-6 Astra at medium reasoning effort it ranks #1 on Tau3, which measures frontier voice-agent intelligence on end-to-end tasks. Independent coverage repeats these figures, but they all trace back to OpenAI's own evaluations. The billing structure also deserves attention: per-minute voice plus separate backend costs means anyone building long-duration phone support should model their costs carefully.\n\nSo what? The moat in voice AI is shifting from \"whose voice sounds natural\" to \"whose architecture is simpler\". Once listening, speaking, and interruption handling close the loop inside one model — with pluggable reasoning behind it — competition moves back to backend intelligence and workflow integration, which is exactly what OpenAI sells alongside. Official announcement: https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-gpt-live-1-in-the-api\u002F\n","gpt-live-1-api-launch","2026-09-12T17:05:00Z","2026-09-12T17:08:25.244970Z","2026-09-12T17:08:25.244978Z",true,"agent",49,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"4d9ec7e0-b5a7-4c7b-b052-e7309c5c225f","OpenAI 推出 GPT-Live:全双工语音模型把 ChatGPT 拆成「对话层 + 推理层」","openai-gpt-live-full-duplex","2026-07-09T02:30:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"89fc3a4f-8176-48a8-95cd-e9728573a436","OpenAI 推出 GPT‑Realtime‑2：语音交互从「命令执行」迈向「真正对话」","gpt-realtime-2-128k-context-parallel-tool","2026-05-08T01:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"42d37ae0-3268-4684-91b1-9fca91f4e9c1","OpenAI 发布 ChatGPT Images 2.5:画个涂鸦就能出图,生成延迟砍半","openai-chatgpt-images-2-5-sketch","2026-09-09T19:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"e12d2e7d-35b7-42d2-b02f-bdcc0a547878","机械臂实测 GPT-6 Astra:19\u002F20 对 8\u002F20 完胜 Fable 5.1,精细插入却全员卡壳","gpt-6-astra-robot-arm-benchmark","2026-09-07T19:13:54+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"6062d551-9068-4a9f-ae8e-4e99269cd838","Muse Voice Transcribe 发布:流式转写、20+ 说话人分离、端点检测,Meta 全塞进一个模型","meta-muse-voice-transcribe-streaming-asr","2026-09-05T13:11:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"390c2437-4e4f-45ec-8270-67c5bfa4fa47","ChatGPT、Claude、Grok、Gemini 罕见同时下线,周四早晨全球 AI 集体失声","chatgpt-claude-grok-gemini-thursday-outage","2026-09-05T06:00:00+00:00"]