OpenAI WebRTC Audio Session, now with document context OpenAI WebRTC 音频会话，现支持文档上下文

So OpenAI has a new realtime voice model that supposedly packs "GPT-5-class reasoning," but if you want to actually use it with your own documents in a conversational audio session, your best bet right now isn't the ChatGPT app you pay for—it's a solo developer's browser playground. Let that sink in.

The developer behind this WebRTC audio tool first built it back in December 2024 when OpenAI's realtime API was still fresh. Now they've updated it with GPT-Realtime-2, a model OpenAI has been quietly hyping as their first voice-first system with serious reasoning chops. It has a knowledge cutoff of September 2024, which already raises eyebrows—by the time most people actually get their hands on these capabilities, that cutoff will feel ancient. The AI industry's favorite trick: announce something impressive, then let the hype cycle do the work while the actual product trickles out over months.

Here's the part that should bother anyone paying attention: this model still hasn't shown up in ChatGPT's iPhone app. OpenAI ships a developer tool, an independent tinkerer builds something useful with it, and meanwhile their flagship consumer product—the one with hundreds of millions of users—is still running last generation's voice capabilities. This isn't a minor oversight. It's a pattern. OpenAI keeps launching capabilities in API-first, scattered fashion while their consumer app lags behind like it's being maintained by a separate, underfunded team. Maybe it is.

The actual feature here—pasting in a chunk of text and having a voice conversation about it—is more interesting than OpenAI seems to realize. Imagine feeding it a contract and asking questions out loud while you're making coffee. Or dumping in meeting notes and having a back-and-forth about action items without staring at a screen. This is the kind of ambient, eyes-free computing interaction that tech companies have been promising since the early Siri days but never quite delivered. The technology is finally here. The packaging is nowhere to be found.

What gets me is the disconnect between what OpenAI demos on stage and what actually lands in your hands. They'll show off a breathtaking realtime conversation at a launch event, complete with dramatic pauses and emotional inflection. Then you open the app and get a voice assistant that still occasionally hallucinates your question or cuts out mid-sentence over spotty WiFi. The gap between prototype theater and production reality at OpenAI has become a chasm.

And let's talk about "GPT-5-class reasoning" as a marketing phrase. What does that even mean in the context of a voice model? Reasoning about what, exactly? The model still has a hard knowledge cutoff. It can't browse the web in real-time during these audio sessions. So we're talking about reasoning over whatever document you paste in, plus whatever it memorized before September. Calling that "GPT-5-class" feels like inflationary branding—the kind of claim that sounds impressive until you ask three follow-up questions.

The solo developer who built this WebRTC playground deserves credit for proving the concept works. Browser-based audio AI conversation with document context is genuinely useful. It's also the kind of thing that should be a standard feature in every major AI product by now. The fact that it exists primarily as someone's side project is an indictment of how slowly the big players move once they've secured your subscription dollars.

OpenAI's real problem isn't technical anymore. It's product discipline. They have incredible models, a massive user base, and a brand that still commands attention. What they don't have is a coherent strategy for getting their best capabilities into the hands of the people who'd actually use them. Instead, we get a patchwork of API updates, developer previews, and consumer app features that seem to follow no particular timeline or priority.

The future of AI interaction is almost certainly voice-first and context-aware. This little WebRTC tool shows exactly why. Talking to your documents feels natural in a way that typing prompts never will. But natural doesn't mean accessible—not yet. Not when the best implementation lives at a URL you have to know about, not in the app everyone already has installed.

Someone at OpenAI should be embarrassed. Someone else should be hiring this developer.

OpenAI推出了一款全新的实时语音模型，据称具备"GPT-5级别的推理能力"。但如果你真想在对话音频场景中使用自己的文档，目前最佳选择并非你付费使用的ChatGPT应用——而是一个独立开发者打造的浏览器端交互平台。请细品其中意味。

OpenAI这款新实时语音模型号称搭载"GPT-5级推理能力"，但当用户试图在对话音频场景中接入自有文档时，目前最实用的解决方案竟不是付费版ChatGPT应用，而是一个独立开发者制作的浏览器交互平台。这个事实值得深思。

该WebRTC音频工具的开发者最初在2024年12月（OpenAI实时API刚发布时）构建了这个项目，近期已升级至GPT-Realtime-2模型。OpenAI一直在低调宣传这是其首款"以语音为先、兼具深度推理能力"的系统，其知识截止日期为2024年9月——这已引发质疑：当普通用户真正接触到这些功能时，该时间点将显得陈旧不堪。这正是AI行业惯用的伎俩：先高调宣布突破性成果，再让市场热度自然发酵，而实际产品可能历经数月才逐步落地。

以下现象应引起所有观察者的警觉：该模型至今仍未登陆ChatGPT的iOS应用。OpenAI发布开发者工具后，独立开发者利用其构建出实用产品，而其拥有数亿用户的旗舰消费级应用，却仍在运行上一代语音能力。这并非微小疏漏，而是某种模式：OpenAI持续以"API优先、零散发布"的方式推出新功能，其消费级应用却如同由资源匮乏的独立团队维护般落后。或许事实正是如此。

真正值得关注的是这项功能本身——粘贴一段文本并围绕其展开语音对话——其趣味性远超OpenAI官方认知。想象一下：煮咖啡时对着合同直接语音提问，或导入会议纪要无需盯屏就能讨论行动项。这种环境感知式、解放双眼的交互体验，正是科技企业自初代Siri以来反复承诺却始终未能实现的场景。技术层面终于成熟，但产品包装却完全缺席。

最令人费解的是OpenAI演示成果与用户实际体验间的巨大落差。