[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-grok-voice-3-s2s-index-top":3,"topics-all":38,"news-related-ff71db84-ec35-4c47-904f-f0f75b3fa0c9":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},"ff71db84-ec35-4c47-904f-f0f75b3fa0c9","Grok Voice 3.0 拿下 S2S 榜首:AA 86.9 分","xAI 10 月 14 日发布 Grok Voice 3.0。AA S2S Index 86.9 超 GPT-Live-1\u002FGemini 3.8 Live,Big Bench Audio 98.5%,全双工 97%,$0.08\u002Fmin、延迟 \u003C1 秒,Tesla 与 Starlink 日均百万级对话。","## 一个让客服电话\"闭嘴\"的发布\n\n10 月 14 日,xAI 正式推出 Grok Voice 3.0——按官方说法,这是当前\"全球最智能的语音转语音模型\"。这套模型已经在 Tesla 车机、Starlink 客服系统里每天承接百万级真实对话。Starlink 客服中心用它每天接 2 万通多语种电话,**约 70% 全程无人工介入就解决了,销售电话里每 4 通就有 **一通在挂断前完成下单。\n\n数字之外,这轮升级最值得说的是定位:Grok Voice 3.0 不是在追一个\"更像人\"的语音助手,而是在**抢\"Agent 时代的对话入口\"**。Voice API 文档已经把它和 Grok Bot、Tesla Connectors 这些可执行任务的 Agent 框架绑在一起——你说一句话,它能调用工具、完成客服工单、甚至下单。这就是它和 GPT-Live-1、Gemini 3.8 Live 这种\"会聊天\"的语音模型之间最大的分野。\n\n## Artificial Analysis 怎么说:S2S Index 86.9,五项核心指标全部第一\n\nArtificial Analysis 在 10 月 13 日发布的独立测评里,Grok Voice 3.0 在 Speech-to-Speech Index 上拿到 **86.9 分**,直接压过 GPT-Live-1 (Astra, Medium) 的 83.2、Gemini 3.8 Live (Extended Thinking, High) 的 82.6,以及自家上一代 Grok Voice 2.0 的 81.3。\n\n把指数拆开看,每一项都是第一名:\n\n| 指标 | Grok Voice 3.0 | GPT-Live-1 | Gemini 3.8 Live |\n|------|----------------|------------|-----------------|\n| S2S Index(综合) | **86.9** | 83.2 | 82.6 |\n| Big Bench Audio(语音推理) | **98.5%** | 90.1% | 97.7% |\n| Task Success Rate(任务完成率) | **94.8%** | 87.4% | 89.1% |\n| Full Duplex Bench(全双工对话) | **97.0%** | 94.9% | 91.9% |\n| Speech Agent Arena Elo | **1070** | 1048 | 990 |\n\n全双工这块,Grok Voice 3.0 在处理停顿、抢话、自然交互上的得分是 97%,比 Gemini 高出 5 个百分点——这是\"会不会打断用户\"、\"会不会在用户停顿时傻住\"这类体验级指标。\n\n## 速度与成本:Agent 落地的两条硬指标\n\n语音模型能不能上生产,只看准确率是不够的。延迟和单任务成本决定它值不值得替换人工。\n\nArtificial Analysis 的 τ-Voice 测试(模拟各种支持电话场景、口音、背景噪音)给 Grok Voice 3.0 打了 **71.6% 的成功率**——略低于 GPT-Live-1 的 74.5%,但**首音频延迟比 GPT-Live-1 低 40%**。在对话式 Agent 场景里,延迟超过 1.5 秒就开始让用户怀疑\"对面是不是真人\",所以这 40% 不是参数游戏,是用户体验的根本性差别。\n\n成本也压得很狠。完成同一组任务,Grok Voice 3.0 单任务成本 **$0.025**,比 GPT-Live-1 的 $0.030 低 17%,比 GPT-Realtime-2.1 的 $0.055 直接砍掉一半多。定价上,Voice Agent API 维持 **$0.08\u002F分钟($4.80\u002F小时)** 与上一代同价,所有用 `grok-voice-latest` 切到新版的 Agent 自动无缝升级。\n\n## 企业语音 Agent 的实战分:EVA-Bench 拿下 74%\n\nServiceNow 维护的 EVA-Bench 是当前少数专门评估\"企业语音 Agent\"的开放基准,200 多个真实场景、跨多语种。Grok Voice 3.0 在准确率上拿了 74%,是当前所有参与模型里的第一。ServiceNow 高级产品总监 Nitin Aggarwal 在引用里给了一句相当重的评语:\"Grok Voice 3.0 在 EVA-Bench 上划出一条新帕累托前沿——把顶级准确率和响应式对话合在一起,这是企业客服一线 AI 期待的新标准。\"\n\n## 所以呢:语音模型终于从\"玩具\"走向\"产线\"\n\n过去两年,语音模型的故事都是\"更像人\"\"更自然\"\"更有温度\"。但要真正替代人工,光像人没用——必须在成功率、延迟、成本上证明它比请一个人便宜、比请一个人快。Grok Voice 3.0 这次走的就是这条路:**S2S Index 第一、Task Success 94.8%、延迟 \u003C1 秒、$0.025\u002F任务**。把这条线的成本打穿,语音 Agent 才能从\"演示片段\"走成\"客服中心的标配\"。\n\n马斯克的赌注很清楚——语音是 Agent 时代最自然的入口,**谁先把这个端点做到上生产质量,谁就拿到下一波企业 AI 预算的入场券**。这一次,xAI 拿出了 4 个独立榜单(AA、Big Bench、Full Duplex、EVA-Bench)+ 真实部署(Tesla、Starlink)的双证据链,值得当回事看。\n\n参考来源:[xAI 官方公告 Grok Voice 3.0](https:\u002F\u002Fx.ai\u002Fnews\u002Fgrok-voice-3) 及独立第三方报道 [SiliconANGLE Grok Voice 3.0 outscores OpenAI and Google](https:\u002F\u002Fsiliconangle.com\u002F2026\u002F10\u002F14\u002Fgrok-voice-3-0-from-spacexsi-outscores-openai-and-google-on-speech-to-speech-index\u002F)。","https:\u002F\u002Fx.ai\u002Fnews\u002Fgrok-voice-3","b82e17a3-1dbd-4b5d-88dc-9f518f917cc0",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"8e273840-b3aa-494e-9132-998a7e26aa78","xai",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"1e6a9018-383f-480b-9dce-3ff3a9ff2908","en","Grok Voice 3.0 tops S2S Index: AA 86.9, under one second","xAI released Grok Voice 3.0 on Oct 14. Artificial Analysis rated it #1 on the Speech-to-Speech Index at 86.9 — ahead of GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6) — with 98.5% on Big Bench Audio, 97% on Full Duplex Bench, $0.08\u002Fmin pricing, sub-second latency, and millions of daily calls already in production across Tesla and Starlink.","## A release that gets customer-support lines to shut up\n\nOn October 14, xAI officially rolled out Grok Voice 3.0 — what the company calls \"the world's most intelligent speech-to-speech model.\" This thing is already live in production: it handles millions of conversations a day across Tesla vehicles and the Starlink call centers. The Starlink deployment alone takes 20,000 multi-language support calls per day; **roughly 70% resolve without any human in the loop, and one in four sales calls ends in a purchase before the caller hangs up**.\n\nUnderneath the headline, the more interesting positioning is this: Grok Voice 3.0 isn't chasing \"sounds more human\" — it's going after **the conversational entry point of the Agent era**. The Voice API documentation already wires it into Grok Bot, Tesla Connectors and other tool-calling agent frameworks. You say something; it can invoke tools, finish a support ticket, even place an order. That's the gap between it and the \"talks nicely\" voice models like GPT-Live-1 or Gemini 3.8 Live.\n\n## What Artificial Analysis said: S2S Index 86.9, all five core metrics #1\n\nIn their independent test dated October 13, Artificial Analysis gave Grok Voice 3.0 a **Speech-to-Speech Index score of 86.9** — beating GPT-Live-1 (Astra, Medium) at 83.2, Gemini 3.8 Live (Extended Thinking, High) at 82.6, and its own predecessor Grok Voice 2.0 at 81.3.\n\nPull the index apart, and it's first place across the board:\n\n| Metric | Grok Voice 3.0 | GPT-Live-1 | Gemini 3.8 Live |\n|--------|----------------|------------|-----------------|\n| S2S Index (overall) | **86.9** | 83.2 | 82.6 |\n| Big Bench Audio (speech reasoning) | **98.5%** | 90.1% | 97.7% |\n| Task Success Rate | **94.8%** | 87.4% | 89.1% |\n| Full Duplex Bench | **97.0%** | 94.9% | 91.9% |\n| Speech Agent Arena Elo | **1070** | 1048 | 990 |\n\nFull Duplex is where the experience lives — handling pauses, turn-taking, not freezing when the user interrupts. Grok Voice 3.0 scores 97% there, a five-point gap over Gemini on the same benchmark.\n\n## Speed and cost — the two hard metrics for production Agent deployment\n\nAccuracy alone doesn't put a model into a contact center. Latency and per-task cost decide whether it's worth replacing a human.\n\nOn Artificial Analysis's τ-Voice test — which simulates support calls with diverse accents, background noise, and interruptions — Grok Voice 3.0 lands at **71.6% agent success rate**, slightly behind GPT-Live-1's 74.5%, but with **40% lower time-to-first-audio latency**. In conversational Agent contexts, anything over 1.5 seconds starts making the caller doubt the agent is real. That 40% is not a parameter game — it's the difference between \"useful\" and \"abandoned by frustrated users.\"\n\nCost is squeezed tight too. To finish the same set of tasks, Grok Voice 3.0 costs **$0.025 per task**, 17% under GPT-Live-1's $0.030 and well below GPT-Realtime-2.1's $0.055 — a 55% cut. API pricing stays at **$0.08\u002Fmin ($4.80\u002Fhr)**, unchanged from Grok Voice 2.0, and every agent on `grok-voice-latest` has already been upgraded automatically with no changes required.\n\n## EVA-Bench: 74% on enterprise voice Agent\n\nEVA-Bench, the open benchmark maintained by ServiceNow, is one of the few evaluations built specifically for \"enterprise voice Agent\" scenarios — 200+ realistic situations across multiple languages. Grok Voice 3.0 hit **74% accuracy** there, the top score of any model on the leaderboard. ServiceNow's Senior Director of PM for Agentic Evals and RAI, Nitin Aggarwal, put a strong line in the official quote: \"Grok Voice 3.0 sets a new Pareto frontier on EVA-Bench — pairing top accuracy with a natural, responsive conversation, setting a new standard for what enterprises can expect from AI on the front line of customer service.\"\n\n## So what: voice models finally cross from \"toy\" to \"production line\"\n\nThe last two years of voice model stories were all about \"more natural,\" \"more human.\" But to actually replace human agents, sounding human isn't enough — you have to prove on success rate, latency, and cost that it's cheaper and faster than hiring a person. Grok Voice 3.0 is the first one to walk it end-to-end: **S2S Index first, Task Success 94.8%, latency under one second, $0.025 per task**. Break that cost line, and voice Agents stop being demo material and become the standard kit of every contact center.\n\nMusk's bet is obvious: voice is the most natural entry point of the Agent era — **whoever that has a production-quality endpoint first gets the entry ticket to the next wave of enterprise AI budgets**. This time, xAI shipped four independent leaderboards (AA, Big Bench, Full Duplex, EVA-Bench) plus real production deployments (Tesla, Starlink) as the dual evidence chain, and it's worth taking seriously.\n\nReference: [xAI official announcement — Grok Voice 3.0](https:\u002F\u002Fx.ai\u002Fnews\u002Fgrok-voice-3), with independent third-party reporting from [SiliconANGLE](https:\u002F\u002Fsiliconangle.com\u002F2026\u002F10\u002F14\u002Fgrok-voice-3-0-from-spacexsi-outscores-openai-and-google-on-speech-to-speech-index\u002F).","grok-voice-3-s2s-index-top","2026-10-15T03:30:00Z","2026-10-15T07:08:05.734336Z","2026-10-15T07:08:05.734346Z",true,"agent",178,[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},"b44fb160-9027-44cd-b355-a66869004abc","Nemotron 3.5 Super VL 开源:120B 总参只激活 12B","nemotron-3-5-super-vl-open-120b-moe","2026-10-20T17:11:56+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"faf5c85e-ed83-432a-8b55-80c9356862c5","腾讯优图开源5B全模态解析:文档榜96.96超GPT-5.2","youtu-parsing-omni-5b","2026-10-19T19:09:27+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"ac666c4e-4756-43d8-840a-4eacbf88f804","Meta 开源 816B MoE 模型 Muse Spark Open,跑在两台 Mac Studio 上","meta-muse-spark-open-816b-mac-studio","2026-10-17T03:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"1b6ec126-cba1-47da-870b-da19cc523763","Meta 开源 816B MoE 模型 Muse Spark Open:1M 上下文,两台 Mac Studio 跑得动","meta-muse-spark-open-816b-moe-1m-mac-studio","2026-10-15T11:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"5d4acf64-67d8-404b-9d3f-cdf7761fb8ac","小鹏开源 GroundingPI:4B 视觉 grounding 模型自评超 GPT-6 Astra","groundingpi-4b-visual-grounding-foundation-model","2026-10-13T17:10:26+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"bf9bbc74-5720-4829-842e-c3a583a87a8b","Qwen3.8 全模态降价九成逼 Gemini 跟牌","qwen3-8-omni-flash-price-cut","2026-10-04T02:00:00+00:00"]