Retell AI（在新窗口中打开） 是一个面向联络中心运营的语音 AI 平台，帮助企业构建能够回答问题、采取行动，并在需要时将呼叫者转接给真人的智能体。其平台目前服务超过 4,000 家企业，月度通话量从 2026 年初的 3,000 万次增长到 9 月的超过 1 亿次。
随着 Retell 为更多具有复杂工作流的企业提供服务，它使用 OpenAI 模型来保持高通话质量和可靠性。如今，OpenAI 模型为 Retell 上 80% 的通话生成回复，客户实现超过 80% 的解决率。Retell 在 OpenAI 模型上消耗的 token 也已超过 1 万亿，反映了这种采用的规模。
Retell 的工程师在整个开发过程中使用 Codex，从研究不熟悉的代码到构建功能和调试问题，并使用 GPT‑6 Astra 测试前端界面。
“我们选择模型的方式就像客户选择供应商一样：看真实通话，而不是演示。OpenAI 一直在赢得它的位置。”
——Bing Wu，Retell AI 联合创始人兼 CEO
为实时通话需求选择 OpenAI 模型
企业语音智能体需要遵循公司政策并保持在指定角色范围内，即使对话发生变化。Retell 构建了 vCX 基准，使用来自真实生产场景的困难多轮对话来比较数十个模型。
它测试模型遵循指令、基于给定信息作答、调用正确工具以及使用自然对话语言的能力。在 Retell 2026 年 10 月 13 日的基准中，GPT‑6 Astra 在指令遵循方面以 89.7% 排名第一。
Retell 还分享了 GPT‑6 Astra 模型发布前进行的评估，结果显示 GPT‑5.5 在启用推理时，在非幻觉率（88.0%）和工具调用正确性（88.3%）两项指标上均排名最高，在这两项指标上都优于 Gemini‑3.1‑Pro 和 Claude‑4.6。在关闭推理时，它也在非幻觉率上领先。
Retell 表示，选择 OpenAI 模型的新客户会继续使用它们。客户还可以通过 Realtime API 访问 OpenAI Realtime 模型，并将 OpenAI 文本转语音作为语音引擎。Retell 在 OpenAI 新模型发布当天即提供这些模型，并随着模型改进重新运行其基准。
用智能体解决超过 80% 的客户请求
联络中心团队使用 Retell 定义其智能体的业务背景、目标和规则。开发者可以进一步配置工作流和集成。
在通话期间，Retell 的运行框架运行音频模型、跟踪对话，并管理智能体何时说话、倾听或响应打断。OpenAI 模型，包括 GPT‑6.1 Sol、GPT‑6 Astra 和 GPT‑5.5，使用客户的指令和对话上下文来生成回复并请求工具调用。Retell 通过与客户系统的集成执行这些调用，并可将结果写回客户记录。
当需要人工判断时，语音智能体会向代表简要说明背景、回答其后续问题，并在连接客户之前确认他们已准备好。通话结束后，Retell 使用 GPT‑5.6 Terra 分析转录文本并提取结果和其他数据。
通过这种方法，Retell 报告客户解决率超过 80%。在电子公司 Anker，智能体在美国和英国处理一级售后支持，包括订单跟踪、保修检查和故障排除。Anker 报告了 80.4% 的案例解决率和 63 的净推荐值（NPS）。其团队还表示，智能体的语气、节奏和温暖程度达到了他们对人工坐席所期望的标准。
在 Pine Park Health，Retell 智能体处理预约提醒和主动排程，并将结果记录在患者的医疗档案中。该团队回收了相当于 2.1 个全职角色的医疗助理时间，并将其重新用于患者护理。其排程智能体在联系到患者时实现了 55.7% 的预约率。
在通话量增长时保持通话响应性
Retell 平台的白天流量可达其夜间低点的 50 倍以上。为了帮助即使在高峰需求时也保持通话响应性，Retell 使用与 OpenAI 合作的专用服务器，并在其客户之间共享。在 Retell 2026 年 10 月 13 日的基准中，GPT‑6 Astra 和 GPT‑6 Sol 的首句延迟中位数均为 883 毫秒，GPT‑5.6 Luna 以 497 毫秒排名最高。
OpenAI 的区域 API 端点允许开发者将流量路由到特定地理区域，这有助于 Retell 减少网络传输时间，并满足医疗和金融服务等行业客户的数据驻留要求。
Retell 还在服务的每一层构建冗余。它运行自己的电话技术栈，并有一个第二技术栈准备接管，同时还有语音和模型服务的备份，以帮助在服务故障时保持通话运行。
使用 Codex 研究、构建和调试
随着 Retell 的企业增长带来更多对自定义工作流和集成的需求，其工程师经常需要在不熟悉的代码库部分工作。在编写代码之前，他们给 Codex 一个高层需求。Codex 会调查代码，对照实际数据检查，并解释需要更改什么。一位工程师说，Codex 在这项研究上做得比他更好，而且快得多。
在实现过程中，工程师表示 Codex 经常捕捉到他们未曾考虑到的细节。对于调试，Codex 会将日志和数据库记录与代码交叉参照，以调查正在发生的情况。Retell 的工程师还在与基于 OpenAI 模型构建的客户合作时使用 Codex，以加速诸如编写指导模型在客户应用中如何回复的指令等任务。
为了测试其 Web 界面的更改，Retell 依赖 GPT‑6 Astra 的计算机使用能力。当工程师打开一个前端拉取请求时，GPT‑6 Astra 会打开界面、点击浏览、测试流程、尝试破坏它们，并报告发现的问题。Retell 表示，早期模型无法可靠地完成这项工作。
Retell 继续招聘，但其工程团队无需以与业务相同的速度增长。Retell 报告每位员工 160 万美元的年化收入。
“通过使用 Codex，工程团队拥有一个虚拟队友，可以在任何类型的困难任务上与他们配对。”
——Bing Wu，Retell AI 联合创始人兼首席执行官
让客户的专业知识发挥作用
每家企业都清楚自己希望如何服务客户。Retell 的目标是让最接近这项工作的人员直接运用他们的专业知识，借助先进的语音 AI 和一个让他们能够构建、测试和改进自己代理的平台。接下来，Retell 计划扩展质量保证和模拟功能，支持更多语言，并开发知识库自动学习能力。
随着 OpenAI 模型不断进步，Retell 期望其客户能够按照自己的方式，为数百万来电者提供世界一流的服务。
“没有人比企业本身更了解自己的客户。我们赋予他们基于这种专业知识采取行动并直接改进 AI 代理的能力。目标是让他们建立一支由内部部署工程师组成的内部团队，全权负责并运营这一切。”Bing 表示。
Retell AI(opens in a new window) is a voice AI platform for contact center operations, helping businesses build agents that can answer questions, take action, and connect a caller with a person when needed. Its platform now serves more than 4,000 businesses, with monthly call volume growing from 30 million in early 2026 to more than 100 million by September.
As Retell serves more enterprises with complex workflows, it uses OpenAI models to keep call quality and reliability high. Today, OpenAI models generate responses for 80% of calls on Retell, with customers achieving resolution rates above 80%. Retell has also surpassed 1 trillion tokens consumed across OpenAI models, reflecting the scale of that adoption.
Retell’s engineers use Codex throughout development, from researching unfamiliar code to building features and debugging issues, and GPT‑6 Astra to test front-end interfaces.
“We choose models the way our customers choose vendors: on real calls, not demos. OpenAI keeps earning its spot.”
—Bing Wu, Co-founder and CEO, Retell AI
Choosing OpenAI models for the demands of a live call
Enterprise voice agents need to follow company policies and stay within their assigned roles, even as conversations change. Retell built its vCX benchmark to compare dozens of models using difficult, multi-turn conversations drawn from real production scenarios.
It tests how well models follow instructions, stay grounded in the information they are given, call the right tools, and use natural conversational language. In Retell’s October 13, 2026 benchmark, GPT‑6 Astra ranked first for instruction following at 89.7%.
Retell also shared evaluations conducted before the launch of the GPT‑6 Astra model that showed GPT‑5.5 ranking highest among models on both non-hallucination rate (88.0%) and tool-call correctness (88.3%) with reasoning enabled, outperforming Gemini‑3.1‑Pro and Claude‑4.6 on both measures. It also led on non-hallucination rate with reasoning off.
Retell says new customers who choose OpenAI models continue using them. Customers also have access to OpenAI Realtime models through the Realtime API and OpenAI text-to-speech as a voice engine. Retell makes new OpenAI models available the day they launch and reruns its benchmark as models improve.
Resolving more than 80% of customer requests with agents
Contact center teams use Retell to define their agents’ business context, goals, and rules. Developers can further configure workflows and integrations.
During a call, Retell’s harness runs the audio models, tracks the conversation, and manages when the agent speaks, listens, or responds to interruptions. OpenAI models, including GPT‑6.1 Sol, GPT‑6 Astra, and GPT‑5.5, use the customer’s instructions and conversation context to generate responses and request tool calls. Retell executes those calls through integrations with the customer’s systems and can write outcomes back to a customer record.
When human judgment is needed, the voice agent briefs the representative with context, answers their follow-up questions, and confirms they are ready before connecting the customer. After a call, Retell uses GPT‑5.6 Terra to analyze the transcript and extract outcomes and other data.
With this approach, Retell reports customer resolution rates above 80%. At electronics company Anker, agents handle Tier‑1 post-sales support in the US and UK, including order tracking, warranty checks, and troubleshooting. Anker reported an 80.4% case resolution rate and a Net Promoter Score (NPS) of 63. Its team also said the agents’ tone, pacing, and warmth met the standard they expected of a human agent.
At Pine Park Health, Retell agents handle appointment reminders and proactive scheduling and record the outcomes in patients’ medical charts. The team recovered medical-assistant time equivalent to 2.1 full-time roles and redirected it to patient care. Its scheduling agent achieved a 55.7% booking rate when it reached a patient.
Keeping calls responsive as volume grows
Daytime traffic on Retell’s platform can reach more than 50 times its overnight low. To help keep calls responsive even at peak demand, Retell uses dedicated servers with OpenAI, shared across its customers. In Retell’s October 13, 2026 benchmark, GPT‑6 Astra and GPT‑6 Sol each recorded median first-sentence latency of 883 milliseconds, with GPT‑5.6 Luna ranking highest at 497 milliseconds.
OpenAI’s regional API endpoints allow developers to route traffic to specific geographic regions, which helps Retell reduce network travel time and meet customers’ data residency requirements in industries such as healthcare and financial services.
Retell also builds redundancy into each layer of the service. It runs its own telephony stack with a second stack ready to take over, along with backups for voice and model services, to help keep calls running when a service fails.
Using Codex to research, build, and debug
As Retell’s enterprise growth brings more requests for custom workflows and integrations, its engineers often need to work in unfamiliar parts of the codebase. Before writing code, they give Codex a high-level requirement. Codex surveys the code, checks it against actual data, and explains what needs to change. One engineer said Codex did this research better and much faster than he could.
During implementation, engineers say Codex often catches details they had not considered. For debugging, Codex cross-references logs and database records with the code to investigate what is happening. Retell’s engineers also use Codex in their work with customers building on OpenAI models to speed up tasks such as writing the instructions that guide how a model responds in a customer’s application.
To test changes to its web interface, Retell relies on GPT‑6 Astra’s computer use. When an engineer opens a front-end pull request, GPT‑6 Astra opens the interface, clicks through it, tests flows, tries to break them, and reports the issues it finds. Retell says earlier models could not perform this work reliably.
Retell continues to hire, but its engineering team hasn’t needed to grow at the same pace as the business. Retell reports $1.6 million in annual revenue run rate per employee.
“By using Codex, the engineering team has a virtual teammate that can pair with them on any kind of difficult task.”
—Bing Wu, Co-founder and CEO, Retell AI
Putting customer expertise to work
Every business knows how it wants to serve its customers. Retell’s goal is to let the people closest to that work apply their expertise directly, using advanced voice AI and a platform that lets them build, test, and improve their own agents. Next, Retell plans to expand quality assurance and simulation, support more languages, and develop knowledge base auto-learning.
As OpenAI models continue to advance, Retell expects its customers to deliver world-class service to millions of callers, on their own terms.
“No one knows a company’s customers better than the company itself. We give them the power to act on that expertise and improve AI agents directly. The goal is for them to build an in-house team of internally deployed engineers who own and run it all,” says Bing.
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文