OpenAI Merges Codex Into ChatGPT With GPT-5.6 Model Upgrade

Codex just got better for developers — OpenAI

OpenAI

OpenAI is integrating its Codex developer tool directly into ChatGPT alongside its new GPT-5.6 Sol model to automate complex multi-step programming tasks across web, mobile, and desktop environments.

AI-assisted summary of Codex just got better for developers — original source. Check the original for full context.

Source published:

Detailed notes

## Background This video was published by OpenAI on YouTube on July 14, 2026. According to the page description, it details what's new in Codex for developers, highlighting features such as GPT‑5.6, Ultra mode, parallel work management, Computer Use updates, inline editing, PR reviews, and Sites.

## TL;DR OpenAI introduced a major update to Codex, integrating it directly into ChatGPT alongside a new work agent and serving over 6 million weekly developers. The release unveils GPT-5.6 Sol and Ultra mode for deep reasoning, alongside automated task division using subagents and background system navigation through Computer Use and 'appshots'. Additional upgrades include one-click app deployment via Sites, cross-project task coordination with git worktrees, mobile workflow capabilities, and fully integrated pull request management.

## Atmosphere & style The presentation is enthusiastic, polished, and developer-focused, conveying a sense of rapid progress and collaborative innovation. The speaker delivers the updates with an energetic yet clear tone, highlighting practical workflows that make complex, multi-agent AI coding feel seamless and interactive. Watching the video feels like an inspiring product showcase that bridges cutting-edge model capabilities with hands-on software development tools.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Over 6 million weekly active developers use OpenAI Codex, which has shipped more than 150 updates in the past two months. - [00:31](https://www.youtube.com/watch?v=eiQgljOrkWU&t=31s) Codex and ChatGPT are merging into a single unified application, placing Codex in a dedicated space alongside the new work agent. - [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) GPT-5.6 Sol model is released for complex problem-solving, supported by an Ultra mode that provides a larger reasoning budget. - [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) Codex automatically breaks down large goals into parallel subagents that can be tracked in a unified management panel. - [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) The Cmd-Cmd shortcut takes an 'appshot' combining visual data and context to enable autonomous desktop and simulator control. - [02:11](https://www.youtube.com/watch?v=eiQgljOrkWU&t=131s) The in-app browser supports visual point-and-click annotations, passkey authentication, and direct inline code diff editing. - [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) Sites is now available to all users, enabling instant full-stack web app deployment with built-in hosting, auth, database, and storage. - [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) Codex can manage complex task graphs, automate issue fixes from Linear, provision isolated git worktrees, and @mention ChatGPT research threads. - [04:57](https://www.youtube.com/watch?v=eiQgljOrkWU&t=297s) The ChatGPT mobile app now supports full Codex task management, detailed git diff reviews, and remote SSH host connectivity. - [05:36](https://www.youtube.com/watch?v=eiQgljOrkWU&t=336s) Pull request workflows are natively integrated into Codex, allowing inline code reviews, PR merging, and automated CI/CD check fixes. - [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) OpenAI announced OpenAI Build Week featuring online programming and nearly 30 in-person global events for developers.

## Detailed walkthrough Here is a detailed, faithful walkthrough of the transcript:

### [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) What's new for developers The speaker introduces this update as a major milestone for OpenAI Codex, expressing enthusiasm to showcase the newest features built for the developer community. Over 6 million developers now use Codex on a weekly basis. Throughout its development, user feedback and ambitious requests—from shipping complex projects to staying organized beyond raw code—have continuously raised the bar and directly shaped the platform. Before diving into the features, the presenter expresses deep gratitude to the user base for their ongoing input.

A key structural change is that Codex and ChatGPT are merging into a unified application. For developers, the experience remains familiar, but Codex now resides within ChatGPT in its own dedicated space next to the new "work agent." The team has maintained a rapid shipment velocity, releasing over 150 updates in the past two months alone, with many more on the horizon.

- Over 6 million weekly active developers use Codex. - More than 150 updates shipped across the preceding two months. - Integration brings Codex directly into the main ChatGPT application interface next to the work agent.

---

### [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) GPT-5.6 Sol and Ultra mode The release introduces **GPT-5.6 Sol**, OpenAI's latest frontier model made available to all users. Sol is engineered to work through complex, high-difficulty problems over extended periods. Alongside Sol, developers can enable the new **Ultra** mode. When combined with the Goal command, Ultra equips Sol with a significantly larger reasoning budget to tackle challenging, long-deferred projects. However, the speaker cautions that operating in Ultra mode will consume token limits at a noticeably faster rate.

Additionally, Codex now features automated task division. Rather than requiring manual prompt breakdown, Codex automatically splits large goals across multiple subagents. Users can inspect subagents individually or open a unified management panel to watch the entire team working in parallel toward a single goal.

- GPT-5.6 Sol provides sustained reasoning capabilities for complex software tasks. - Ultra mode allocates a larger compute/reasoning budget, utilizing token limits faster. - Subagents automatically partition tasks to run in parallel without manual user orchestration.

---

### [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) Computer Use and appshots The presenter highlights GPT-5.6's computer and browser use capabilities as perhaps the most exciting feature of the release, noting its dramatic speed improvements. Demonstrating on an iPhone application inside the macOS Simulator, the presenter captures an **"appshot"** by pressing `Cmd-Cmd` on Mac. An appshot functions as an intelligent screenshot, supplying Codex with both visual screen data and underlying app context.

Upon receiving the appshot with the prompt to "navigate through every screen and capture the App Store screenshots in English and French," Codex directly controls the macOS Simulator. The presenter points out that the visible cursor moving on screen is not human input, but Codex autonomously navigating the app in the background while the user continues other desktop tasks uninterrupted.

- `Cmd-Cmd` hotkey on macOS takes an "appshot" (a screenshot packaged with visual and contextual application metadata). - Codex directly interacts with desktop elements, such as clicking through an iOS Simulator, in background threads.

---

### [02:11](https://www.youtube.com/watch?v=eiQgljOrkWU&t=131s) Browser Use and inline editing Shifting to web development, the video demonstrates a simple paper plane game built with GPT-5.6. Opening the game directly in Codex's built-in browser, developers can play and inspect the output. The updated browser allows users to point and click directly on visual elements to annotate specific areas, giving Codex visual context to apply targeted modifications. The browser engine has also been updated to support applications requiring authentication, including full support for **passkeys**.

In addition to visual annotations, Codex improves code editing workflows. Developers can inspect code diffs and make precise direct edits inline. If a developer knows the exact fix, they can type the line of code themselves; if they only know the desired behavior, they can prompt Codex to handle the implementation. This dual visual-and-code approach creates an environment that feels "much closer to actually pairing with the model."

- In-app browser supports visual point-and-click annotations for localized code modifications. - Browser includes support for authenticated sessions and passkeys. - Inline editing allows manual code fixes inside diff views or prompt-driven changes directly at the source line.

---

### [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) Publishing with Sites To solve the deployment bottleneck, OpenAI has expanded **Sites**. Initially introduced the previous month for Enterprise customers to deploy internal tools, Sites is now available to all users for any web application built within Codex.

By issuing a simple command like "publish this to Sites," developers can instantly deploy full-stack applications. Sites includes built-in infrastructure out of the box: - Application hosting - Built-in authentication - A persistent database - Integrated file storage

This allows developers to move from initial idea to a live web deployment without manually configuring servers, hosting platforms, or external cloud infrastructure.

---

### [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) Coordinating work across projects As developers manage multiple concurrent projects, Codex expands its orchestration capabilities. Codex can now manage its own internal task graph. In a practical example, the presenter issues a command to: "Find five bugs from Linear, create a separate task and isolated worktree for each one, and pin the most critical ones."

From a single coordinating thread, Codex provisions git worktrees and tracks execution across independent tasks. These individual tasks can reference one another for context. Furthermore, context can be brought in seamlessly from other OpenAI tools: if an idea or technical specification is developed in ChatGPT using Deep Research, the user can `@mention` that specific conversation thread in Codex to import all relevant context without copying and pasting.

- Integration with Linear for automated issue discovery and task creation. - Creation of isolated git worktrees for independent subtasks. - Ability to `@mention` prior ChatGPT and Deep Research threads directly inside Codex tasks.

---

### [04:57](https://www.youtube.com/watch?v=eiQgljOrkWU&t=297s) Codex on mobile The ChatGPT mobile application receives significant upgrades to support developer workflows away from the desktop. The updates center on three main capabilities: 1. **Task Management:** Users can search, create, open, and manage Codex tasks directly from conversation threads on their phone. 2. **Diff & Branch Reviews:** Mobile interface updates allow filtering between staged, unstaged, branch-level, or cross-branch comparisons. 3. **SSH Connectivity:** Users can establish direct connections to remote SSH hosts (such as cloud dev boxes) and trigger new task executions directly from a mobile device.

These tools allow developers to kick off jobs at a workstation and monitor or assist Codex while on the go.

---

### [05:36](https://www.youtube.com/watch?v=eiQgljOrkWU&t=336s) Pull request workflows On the desktop environment, pull request handling is now deeply integrated into the main Codex interface. Codex operates with awareness of the active git branch for each task. When an open pull request exists for that branch, it is automatically surfaced in the summary pane.

Within the dedicated **Pull Requests** tab, developers can perform complete code reviews without navigating to GitHub.com: - Review code changes and leave inline review comments. - Ask Codex to fix broken CI/CD checks or address review feedback directly in context. - Mark pull requests as ready for review and execute merges directly from within the application.

- Integrates PR management natively to avoid context switching to GitHub.com. - Automatically connects active task branches to their corresponding open PRs. - Allows automated resolution of failing status checks and review comments.

---

### [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) OpenAI Build Week Concluding the presentation, the speaker emphasizes that these feature updates reflect an ongoing collaboration between OpenAI and the developer community. OpenAI invites developers to participate in **OpenAI Build Week**, featuring online programming as well as nearly 30 in-person events around the world.

To end the demonstration, the presenter prompts Codex to wrap up the video. Codex delivers a closing message encouraging developers to continue submitting "your hardest bugs, your boldest ideas, and the work that feels slightly impossible."

- OpenAI Build Week includes online events and nearly 30 in-person global gatherings.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Over 6 million developers now use Codex on a weekly basis.

> [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Codex and ChatGPT are merging into a unified application.

> [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) When combined with the Goal command, Ultra equips Sol with a significantly larger reasoning budget to tackle challenging, long-deferred projects.

> [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) Operating in Ultra mode will consume token limits at a noticeably faster rate.

> [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) An appshot functions as an intelligent screenshot, supplying Codex with both visual screen data and underlying app context.

> [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) Navigate through every screen and capture the App Store screenshots in English and French

> [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) Publish this to Sites

> [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) Find five bugs from Linear, create a separate task and isolated worktree for each one, and pin the most critical ones.

> [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) OpenAI Build Week, featuring online programming as well as nearly 30 in-person events around the world.

> [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) Your hardest bugs, your boldest ideas, and the work that feels slightly impossible.

## Mentioned OpenAI, OpenAI Codex, ChatGPT, GPT-5.6 Sol, GPT-5.6, Ultra mode, Goal command, macOS Simulator, App Store, Sites, Linear, Deep Research, GitHub.com, OpenAI Build Week

[Source](https://www.youtube.com/watch?v=eiQgljOrkWU)

中文简报

OpenAI 将 Codex 整合至 ChatGPT 并升级 GPT-5.6 模型 OpenAI 正在将其 Codex 开发者工具直接集成到 ChatGPT 中,并搭载全新的 GPT-5.6 Sol 模型,以在网页端、移动端和桌面端环境中实现复杂多步骤编程任务的自动化。 OpenAI 对 Codex 进行了重大更新,每周服务超过 600 万名开发者,并将其与新的工作智能体一同直接集成到 ChatGPT 中。此次发布推出了 GPT-5.6 Sol 以及用于深度推理的 Ultra 模式,同时支持通过子智能体自动拆解任务,并借助 Computer Use 和“appshots”进行后台系统导航。其他升级还包括通过 Sites 一键部署应用、利用 git worktrees 协调跨项目任务、支持移动端工作流,以及完全集成的 Pull Request 管理。 OpenAI 宣布 Codex 与 ChatGPT 正式整合至统一应用中,现已有超 600 万周活跃开发者使用,且在过去两个月内推出了逾 150 项更新。 (0:00) 推出全新前沿模型 GPT-5.6 Sol 及其配套的 Ultra 模式,提供更大推理预算并能将庞大任务自动拆解给多个并行子智能体(subagents)协同处理。 (0:46) 升级 Computer Use 功能,通过 Cmd-Cmd 快捷键拍摄“appshot”即可在后台自主操控 macOS 模拟器;内置浏览器亦支持点选视觉批注、通行密钥(Passkey)及行内代码 diff 直接编辑。 (1:28) Sites 现已面向所有用户开放,支持开箱即用的全栈托管、身份验证与数据库存储;同时 Codex 支持多项目协同,可对接 Linear 自动创建独立 git worktrees 并通过 @ 引用 ChatGPT 研究上下文。 (3:17) ChatGPT 移动端新增任务管理、分支 diff 审查及远程 SSH 连接功能,同时桌面端深度集成了 GitHub PR 审查、自动修复 CI 报错与直接合并等工作流。 (4:57)

Reading summary

OpenAI has integrated Codex directly into the ChatGPT interface alongside its new work agent, pairing the structural merge with the launch of GPT-5.6 Sol and expanded autonomous development tools for its base of over 6 million weekly active developers. Advanced Reasoning and Automated Subagent Division (0:46) The update introduces GPT-5.6 Sol, OpenAI's latest frontier model designed to sustain reasoning across complex, high-difficulty software engineering problems over extended periods. Developers can pair Sol with the Goal command to activate Ultra mode, which allocates a substantially larger reasoning and compute budget for challenging projects. However, the presenter cautions that operating in Ultra mode will consume token limits at a noticeably faster pace. Codex also introduces automated task orchestration. Rather than requiring developers to manually partition prompts, Codex automatically divides high-level goals across multiple subagents running in parallel. Users can monitor execution by inspecting subagents individually or through a unified management panel that tracks the coordinated team. Autonomous Desktop Navigation and In-Browser Editing (1:28) The release upgrades Codex's computer and browser control capabilities. On macOS, developers can press the `Cmd-Cmd` shortcut to capture an "appshot," which bundles visual screen data with underlying application metadata. During a demonstration on an iOS Simulator, Codex autonomously navigated through application screens to generate localized App Store screenshots in English and French in the background, without requiring active human cursor control. For web development, Codex features a built-in browser engine that supports authenticated sessions and passkeys. Developers can point and click on visual UI elements to supply targeted context for model-driven adjustments. This visual feedback works alongside an inline code diff editor that allows developers to either hand-type exact code fixes or prompt Codex directly at the specific line. Multi-Task Orchestration and Native Git Workflows (3:51) To coordinate multi-project workflows, Codex now manages an internal task graph and integrates with external project trackers like Linear. In one demonstration, the model queried Linear for bugs, isolated each issue within its own git worktree, and tracked progress across independent tasks from a single thread. Tasks can cross-reference one another, and developers can `@mention` prior ChatGPT and Deep Research threads to import technical specifications without manual copying. Codex also integrates git pull request management directly into its desktop interface, linking tasks to active branches. Within a dedicated Pull Requests tab, developers can inspect diffs, post inline review comments, direct Codex to remediate broken CI/CD checks, and merge branches without navigating to GitHub.com. Full-Stack Publishing and Mobile Remote Control (3:17) OpenAI expanded its Sites deployment platform from an Enterprise-only internal tool to all Codex users. Developers can trigger complete application deployments with commands such as "publish this to Sites." The platform automatically provisions built-in application hosting, authentication, a persistent database, and integrated file storage, removing the need to configure separate cloud infrastructure. Additionally, the ChatGPT mobile application has been updated to support off-desktop engineering workflows. Developers can create and manage Codex tasks, review staged or branch-level code diffs, and connect directly to remote SSH environments, such as cloud dev boxes, to initiate and supervise runs while away from their primary workstations. OpenAI reported shipping more than 150 platform updates over the preceding two months and concluded the release by announcing OpenAI Build Week, which includes online sessions and nearly 30 in-person global events. While the demonstration highlights autonomous execution across local simulators, cloud infrastructure, and git workflows, the presenter acknowledged practical constraints, including accelerated token consumption when running models in Ultra mode.

精炼阅读摘要

OpenAI宣布对Codex进行重大升级并将其整合至ChatGPT主界面,通过引入GPT-5.6 Sol模型与多智能体协作机制,将开发辅助从代码补全拓展为涵盖桌面操作、全栈部署与拉取请求评审的端到端工程平台。 核心模型升级与智能体并行协作 (0:46) Codex引入了最新前沿模型GPT-5.6 Sol,旨在为复杂且长周期的软件工程任务提供持续推理支持。配合Goal指令,开发者可开启全新的Ultra模式,该模式为疑难项目分配更大的计算与推理预算,但演讲者提醒其消耗Token额度的速度会显著加快。此外,Codex具备自动化任务拆解能力,能够自动将高阶目标分发给多个子智能体(subagents)并行处理,用户无需手动拆解提示词,即可通过集中管理面板实时查看团队式协作进度。 在架构与用户基数方面,Codex已正式并入ChatGPT主应用程序,与工作智能体相邻设立专属空间。演讲者公布的数据显示,Codex当前每周活跃开发者数量已超过600万,团队在过去两个月内累计发布了超过150次更新。 桌面与浏览器环境的自主操作 (1:28) GPT-5.6大幅提升了计算机与浏览器操控能力。在macOS环境下,开发者使用Cmd-Cmd快捷键即可生成“appshot”(一种打包了屏幕画面与应用底层上下文元数据的智能快照)。在展示中,Codex接收到针对iOS模拟器的指令后,能够在后台独立移动光标并遍历页面,自主截取英法双语应用商店截图,全程不中断用户的桌面常规操作。 在Web开发场景中,内置浏览器新增了点选交互功能,开发者可以直接点击界面视觉元素进行标注,为Codex提供精确的定位上下文以执行定向修改。该浏览器还集成了通行密钥(passkeys)等会话认证支持。代码编辑方面,Codex在差异(diff)视图中支持内联编辑,开发者既能亲手修改代码行,也可直接在源码位置输入提示词交由模型实现。 Sites即时部署与跨项目工作流编排 (3:17) 针对部署环节的瓶颈,OpenAI将此前仅面向企业客户的Sites功能向所有用户开放。开发者在Codex中仅需输入发布指令,即可一键上线全栈Web应用。Sites开箱即用集成了应用托管、身份验证、持久化数据库和文件存储功能,避免了手动配置外部云服务器与基础设施的繁琐流程。 针对多任务协同,Codex支持管理内部任务图谱。演示中,系统展示了对接Linear问题追踪系统的能力,可自动提取缺陷并为各个子任务分别建立相互隔离的git worktree(工作树),便于并行推进并保持上下文相互引用。同时,用户可以在Codex任务中通过@mention直接调用ChatGPT中Deep Research的研究记录,实现跨会话工程背景信息的无缝导入。 移动端控制与端到端PR工作流 (4:57) ChatGPT移动应用在此次发布中强化了对离职守工作流的适配。开发者可在手机端创建并检索Codex任务、在暂存区和各分支之间筛选比对代码差异,并能建立SSH连接直接操作远程开发机执行作业,使开发者在外出时仍能监督和推进桌面任务。 在桌面端,拉取请求(Pull Request)深度整合进Codex主界面。系统能够根据当前活跃的git分支自动关联对应PR,开发者无需切换至GitHub网页端,便可在界面内完成行内代码审查、要求Codex针对失败的CI/CD检查自动修复报错,并直接在应用内将PR标记为就绪或执行合并操作。 演讲者以Codex自动生成的寄语为演示收尾,并公布了包含近30场全球线下聚会的OpenAI Build Week活动。整场发布展示了OpenAI将Codex推向完整开发生命周期控制中枢的意图,但Ultra模式带来的更高Token消耗率以及复杂多智能体协同在真实大型工程中的稳定性,仍需开发者在实践中持续评估。

Detailed notes — zh

## Background 该视频由 OpenAI 于 2026 年 8 月 3 日发布在 YouTube 官方频道。视频简介指出,本次更新汇集了大量面向开发者的重磅特性,涵盖 GPT-5.6、Ultra 模式、并行任务管理、大幅升级的 Computer Use、行内代码与文档编辑、PR 审查以及 Sites 一键部署等,并在片尾预告了即将亮相的未公开功能。 ## TL;DR 视频全面展示了 Codex 与 ChatGPT 深度融合后的最新能力,目前其周活跃开发者已突破 600 万。核心更新包括引入具备极强长程推理能力的 GPT-5.6 Sol 及 Ultra 模式,支持自动派生子智能体团队并行攻坚;推出具备深层应用上下文的 Appshot 与后台无感运行的 Computer Use,并支持内置浏览器即时交互与 Diff 视图双向行内编辑。此外,自带数据库与认证的全栈托管服务 Sites 面向全体开发者开放,同时新增跨项目任务协调、ChatGPT 移动端运维审查与 SSH 直连,以及原生的 GitHub PR 闭环工作流。 ## Atmosphere & style 整场演示节奏利落且极客感十足,主讲人以谦逊真诚的姿态向开发者社区致谢,将各项硬核特性以极具现场感的结对编程视角娓娓道来。从模拟器自主穿梭的丝滑后台操作,到多分支流水线的自动拆解,处处传递着将开发者从机械劳动中彻底解放的务实信念。结尾处 Codex 亲自发声号召大家交付“最棘手的 Bug 与看似不可能的大胆构想”,在实用高效之余碰撞出强烈的未来共创感与技术感染力。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Codex 与 ChatGPT 正式整合并入单一应用,周活跃开发者突破 600 万人,并与全新工作智能体(work agent)协同拥有专属独立工作区。 - [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) 新一代前沿推理模型 GPT-5.6 Sol 面向全员开放,引入 Ultra 模式扩展推理预算,可基于单个目标指令自主拆解并并行调度多个子智能体协同攻坚。 - [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) 全栈托管平台 Sites 正式向所有开发者开放,支持通过单句提示词将 Web 应用秒级发布上线,开箱集成身份认证、持久化数据库和文件存储。 - [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) 跨项目与移动端工作流实现全面闭环,Codex 支持从 Linear 自动分配隔离的 Git Worktree 任务群,并在 ChatGPT 移动端提供远程 SSH 直连与完整 PR 审查合并能力。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Codex 与 ChatGPT 深度合并进单一应用,保持高频迭代,过去两个月内已连续推送超 150 项功能更新,周活跃开发者突破 600 万。 - [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) 全新前沿模型 GPT-5.6 Sol 正式面向全体用户开放,具备强大的长程复杂推理能力,擅长持续攻克高难度工程挑战。 - [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) 系统引入 Ultra 模式以扩展推理预算,与 Goal 指令结合可专注解决顽固核心难题,但 Token 消耗速度也会相应加快。 - [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) Codex 具备自主编排能力,无需人工干预提示即可将总目标拆解并分发给多个子智能体并发执行,并提供全局面板集中监控协作进度。 - [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) Computer Use 带来智能截图 Appshot(Mac 端双击 Command 触发),不仅获取界面视觉,更能深度抓取应用的上下文环境与结构。 - [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) 智能体实现了完全无感的真后台自动化操作,能自主驱动 iOS 模拟器翻页并捕获多语言素材,全程不抢占物理鼠标光标与用户焦点。 - [02:11](https://www.youtube.com/watch?v=eiQgljOrkWU&t=131s) 应用内原生集成浏览器,兼容 Passkey 免密登录,并提供可视化标注工具让开发者直接在界面元素上点击批注以指导模型重构。 - [02:11](https://www.youtube.com/watch?v=eiQgljOrkWU&t=131s) Diff 视图升级支持双向行内代码编辑,开发者既能直接在对比视图中修改单行代码,也能针对特定行就地向 Codex 提问让其代劳。 - [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) 全栈部署服务 Sites 从企业内测扩展至全体开发者,只需输入“publish this to Sites”即可一键发布任意 Codex 构建的 Web 应用。 - [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) Sites 并非单纯的静态页面托管,其底层原生打包了全栈生产要素,包括自动化托管、用户身份认证、持久化数据库与文件存储。 - [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) Codex 具备跨项目任务自编排能力,可解析 Linear 中的 Bug 列表,自动为每个问题分配隔离的 Git Worktree 并置顶高优先级任务。 - [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) 开发流程打破孤岛限制,开发者可通过 @mention 语法直接在代码区调取引导自 ChatGPT 日常对话或 Deep Research 的调研脉络。 - [04:57](https://www.youtube.com/watch?v=eiQgljOrkWU&t=297s) ChatGPT 移动端全面进化,开发者在手机上即可利用自然语言完成 Codex 任务的全生命周期创建、检索、管理与流转。 - [04:57](https://www.youtube.com/watch?v=eiQgljOrkWU&t=297s) 移动端支持直连远端 SSH 开发机(Dev Box)随时触发构建任务,并配备支持暂存区、未暂存区及跨分支对比的成熟代码审查器。 - [05:36](https://www.youtube.com/watch?v=eiQgljOrkWU&t=336s) 桌面端深度原生集成 Pull Request 工作流,自动关联 Git 分支状态,支持根据 CI 报错与审查反馈就地修复代码并一键 Merge。 - [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) OpenAI 宣布举行 Build Week 活动,将在全球近 30 个线下城市联动线上直播展开,鼓励开发者向 Codex 提交最具挑战性的任务与创意。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) Codex 与 ChatGPT 深度融合及每周更新 对于 Codex 而言,这是一个具有里程碑意义的时刻。主讲人难掩兴奋地向开发者们展示一系列全新特性。如今每周有超过 600 万名开发者在使用 Codex,社区不断分享反馈、展示作品,并为产品设定了极高的标准——从交付野心勃勃的大型项目,到在编写代码之外帮助开发者保持工作井井有条。主讲人特别向所有一路陪伴、塑造 Codex 的开发者表达了诚挚谢意:"You've helped shape Codex at every step, so before I show you anything, thank you." Codex 与 ChatGPT 现已正式合并到同一个应用程序中。对开发者来说,这依然是大家熟悉且热爱的 Codex,但它现在入驻于 ChatGPT 内部,并在全新的“工作智能体(work agent)”旁拥有专属独立工作区。过去两个月内,团队保持着每周高频迭代的节奏,累计推送了超过 150 项更新,而这仅仅只是一个开始。 - 每周活跃开发者数量已突破 600 万人。 - Codex 与 ChatGPT 整合为单一 App,在工作智能体旁拥有专属独立空间。 - 过去两个月内已连续发布超过 150 项功能改进与更新。 --- ### [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) GPT-5.6 Sol、Ultra 模式与多智能体并行协作 作为核心引擎,全新前沿模型 GPT-5.6 Sol 正式面向所有用户开放。该模型具备极强的长程复杂推理能力,能够长时间专注于攻克高难度问题。与此同时,系统还引入了搭配 GPT-5.6 使用的“Ultra 模式”。当 Ultra 模式与 `Goal` 命令结合时,开发者可以将那些一直搁置、极其棘手的核心难题直接交给它处理。Ultra 模式赋予了 Sol 更充裕的推理预算(reasoning budget),不过主讲人也幽默地提醒大家要留心 Token 消耗速度:"Ultra will obviously be using your token limits faster." 在实际执行任务时,Codex 展现出令人惊叹的自主编排能力:它会自动将大目标拆解并分发给多个子智能体(subagents)并行推进,完全不需要人工干预提示。开发者既可以单独追踪某一个子任务的动态,也可以展开全局协作面板,纵览整个智能体团队协同作战的全景,实现“一个目标,全线并行”。 ```mermaid flowchart TD accTitle: Codex 多智能体任务分发架构 accDescr: 展示从输入目标到由 Ultra 模式驱动的子智能体并行执行流程 GoalInput["开发者设定总目标 (Goal 指令)"] UltraEngine["GPT-5.6 Sol (Ultra 模式)<br/>扩展推理预算"] subgraph SubagentsGroup ["自动子智能体团队 (Subagents)"] Sub1["子智能体 A<br/>架构与依赖处理"] Sub2["子智能体 B<br/>核心功能实现"] Sub3["子智能体 C<br/>测试与验证"] end CoordPanel["全局协作面板 (Coordinating View)<br/>并行监控与进度推进"] GoalInput --> UltraEngine UltraEngine --> Sub1 UltraEngine --> Sub2 UltraEngine --> Sub3 Sub1 --> CoordPanel Sub2 --> CoordPanel Sub3 --> CoordPanel ``` - **GPT-5.6 Sol**:最新前沿模型,面向全体用户开放,擅长长时间攻坚难题。 - **Ultra 模式**:大幅提升模型的推理预算,专为高难度长任务设计(Token 消耗速率同步上升)。 - **自主子智能体协作**:无需人工干预即可将任务自动拆解并行化,支持单任务独立跟踪与团队全局看板监控。 --- ### [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) Computer Use 升级与智能截图 Appshot 在本次发布的所有更新中,主讲人最为推崇的是与 GPT-5.6 结合后的 Computer Use 与 Browser Use 能力,其执行速度与准确度均实现了飞跃。现场演示了一个 iPhone 应用开发的真实场景:当开发者需要为应用商店准备素材时,无需再手动截取数十张不同尺寸和语言的图片。 在 Mac 上只需连按两次 Command 键(Cmd-Cmd),即可对 iOS 模拟器截取一张 "Appshot"。Appshot 被形容为一种更聪明的截图("a smarter screenshot")——Codex 不仅能获取当前屏幕的视觉画面,还能完整读取应用的上下文环境与结构。紧接着,主讲人下达指令要求其遍历应用的每一个界面,分别捕获英语和法语版本的 App Store 宣传截图。令人惊艳的是,模拟器中的光标开始自动点击、翻页和导航,所有操作均在后台静默运行,完全不会抢占开发者的鼠标焦点或干扰日常电脑操作。 - **Appshot 交互**:Mac 端双击 Command 快捷键捕获,兼具屏幕视觉信息与深层应用上下文。 - **模拟器完全自动化**:自主控制 iOS 模拟器完成深层次页面跳转、多语言界面切换与截图导出。 - **真后台无感运行**:智能体操作不抢占物理鼠标光标,开发者可在前台继续正常工作。 --- ### [02:11](https://www.youtube.com/watch?v=eiQgljOrkWU&t=131s) 内置浏览器交互与 Diff 视图行内代码编辑 针对 Web 开发,Codex 深度集成了应用内浏览器(in-app browser)。演示中使用 GPT-5.6 构建了一个简易的纸飞机游戏,开发者可直接在内置浏览器中试玩,实时观察运行效果。如果发现界面有需要调整的地方,可以直接利用可视化标注工具“指哪打哪”,点击画面中的特定元素进行批注,Codex 会借助视觉上下文精准理解意图并完成修改。同时,浏览器还强化了登录场景支持,原生兼容 Passkey 免密认证。 对于代码细节的把控,Codex 升级了 Diff 代码审查视图。过去开发者只能在对比视图中查看改动,现在则可以直接在 Diff 视图中进行精确的“行内代码编辑(inline edits)”。如果你明确知道如何修改,可以直接上手改动单行代码;如果你只清楚预期表现而不确定底层逻辑,也可以直接在该行向 Codex 提问让其代劳。这种将视觉点选与行内代码掌控相融合的体验,主讲人感叹道,这真正带来了与 AI 模型并肩结对编程(Pair Programming)的沉浸感。 - **应用内原生浏览器**:支持即时预览、多步骤交互测试与 Passkey 登录鉴权。 - **视觉化指针标注**:直接点击预览画面添加修改批注,为模型提供直观的空间视觉上下文。 - **Diff 视图双向行内编辑**:既支持人工直接修改特定行,也支持以当前代码行为基准下发指令让模型重构。 --- ### [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) 面向全员开放的 Sites:一键全栈部署上线 当一个应用开发完成,开发者面临的下一个问题通常是:“我该如何把它发布到互联网上?”("how do I put it on the internet?")。上个月 OpenAI 面向企业客户推出了 Sites,支持团队内部共享原型与工具;而在本次发布中,Sites 正式面向所有 Codex 用户开放,适用于使用 Codex 构建的任何 Web 应用。 部署流程被简化到极致:开发者只需对 Codex 发送一句简单的指令——"publish this to Sites"(发布到 Sites)。该平台底层并非单纯的静态页面托管,而是高度集成了自动化托管、用户身份认证(Authentication)、持久化数据库(Persistent Database)以及文件存储(File Storage)。开发者无需为配置服务器、云凭证或数据库运维分心,几秒钟内即可实现从单纯的代码逻辑到具备生产要素的全栈应用一键上线。 - **开放范围**:从企业专享正式普及至全体开发者,覆盖任意 Codex 构建的 Web 应用。 - **全栈开箱即用**:内建托管环境、身份认证系统、持久化数据库和文件存储基础设施。 - **极简部署指令**:单句提示词即可完成自动化构建、环境初始化与公网部署。 --- ### [03:51](https://www.youtube.com/watch?v=eiQgljOrkWU&t=231s) 跨项目任务协调与上下文无缝打通 随着开发者将越来越多跨项目的繁杂事务委托给 Codex,如何统筹全局成为核心痛点。Codex 现已具备自主管理自身任务(Codex Tasks)的高阶能力。主讲人现场演示了一个多项目统筹的经典场景:他在统一的协调会话线程中输入指令——从 Linear 中检索 5 个 Bug,为每一个 Bug 分配独立的任务与隔离的 Git Worktree,并将最关键的问题置顶置优。 ```mermaid flowchart LR accTitle: 跨平台多任务自动编排与隔离流水线 accDescr: 展示从 Linear 问题抓取到创建独立工作区及任务双向关联的流转图 subgraph Source ["输入源"] Linear["Linear Bug 列表"] ChatContext["ChatGPT 会话 /<br/>Deep Research 探索"] end subgraph Coordinator ["Codex 协调线程"] Orchestrator["任务智能编排器"] end subgraph Workspaces ["隔离工作空间 (Git Worktrees)"] T1["Bug 1: 独立 Worktree (高优先级置顶)"] T2["Bug 2: 独立 Worktree"] T3["Bug 3: 独立 Worktree"] end Linear --> Orchestrator ChatContext -.->|"@mention 引入上下文"| Orchestrator Orchestrator --> T1 Orchestrator --> T2 Orchestrator --> T3 T1 <--> T2 ``` 顷刻之间,主线程便将 5 项任务全部拉起并独立推进,各任务之间还支持相互引用与交叉依赖。更具突破性的是,研发流程不再强制始于 Codex 内部:如果某个灵感诞生于 ChatGPT 的日常对话,或者源自深度研究功能(Deep Research)的调研结果,开发者现在只需在 Codex 中输入 `@` 提及对应的历史会话,就能直接将完整的思考脉络与背景资料无损注入代码空间,彻底告别繁琐的手动复制粘贴。 - **自我任务管理**:能够解析外部项目跟踪系统(如 Linear),自动划分独立任务并建立互不干扰的 Git Worktree。 - **跨任务相互关联**:各子任务间可双向链接、置顶关键事项并在主协调面板集中汇报。 - **全局上下文打通**:支持 `@mention` 引用 ChatGPT 历史讨论或 Deep Research 调研成果,实现无缝衔接。 --- ### [04:57](https://www.youtube.com/watch?v=eiQgljOrkWU&t=297s) ChatGPT 移动端赋能:随时随地管理任务与 SSH 连接 针对开发者离开工位、无法守在电脑前的情景,ChatGPT 移动端迎来了一整套重磅升级。主讲人重点演示了三大核心能力: 1. **任务全生命周期管理**:在手机上可以直接通过自然语言对话创建、检索、打开并管理正在运行的 Codex 任务。 2. **多维度代码审查(Diff Review)**:移动端的代码差异查看器极为成熟,允许开发者在暂存区(staged)、未暂存区(unstaged)、具体分支以及跨分支对比(compare branches)之间自如切换。 3. **直连 SSH 主机(Dev Box)**:开发者能够在手机端直接绑定远端 SSH 开发机,随时随地在远程环境中触发并启动全新的编译或排障任务。 这套移动端能力打通了桌面与户外的界限。主讲人表示自己非常享受这种自由的工作模式:在办公室的工位上启动一个项目,在通勤途中通过手机端持续跟进进展,仅在 Codex 真正遇到阻塞或需要最终确认时介入指导。 - **手机端任务枢纽**:直接在移动端会话中对任务进行增删查改和状态流转。 - **移动级代码审查**:支持 staged、unstaged 及分支间对比的高级筛选视图。 - **远端 SSH 接入**:可在手机上与远程开发机(Dev Box)建立安全通信并直接调度任务。 --- ### [05:36](https://www.youtube.com/watch?v=eiQgljOrkWU&t=336s) 原生 Pull Request 工作流与 OpenAI Build Week 回到桌面端,日常开发流迎来了最后一块拼图——Codex 实现了上下文感知的原生 Pull Request(PR)集成,彻底免去了在本地编辑器与 GitHub.com 之间来回切换的割裂感。Codex 能够自动感知当前任务所关联的 Git 分支;当该分支在远端存在打开的 PR 时,摘要面板会自动将其浮现出来。 如果持续集成(CI)检查失败,或者同事提出了审查意见(review comments),开发者可以直接在上下文面板中要求 Codex 立即针对该反馈修复代码。此外,全新的“Pull Requests”标签页提供了完整的代码审查界面:开发者可以逐行查阅代码变动、撰写行内评审意见、将自己的 PR 标记为 Ready,并在审查通过后直接就地完成 Merge 合并。 在演示的尾声,主讲人强调这些新特性只是近期发布内容的一部分,而团队的核心信念始终未变——那就是与开发者并肩同行。随后,主讲人邀请全球开发者参与即将到来的“OpenAI Build Week”,该活动将通过线上直播以及在全球近 30 个城市举办的线下活动同步展开。视频最后出现了一段俏皮的彩蛋:主讲人让 Codex 自己来为 Build Week 预热并收尾,Codex 以充满激情的口吻向开发者发出号召:"As always, keep giving us your hardest bugs, your boldest ideas, and the work that feels slightly impossible. Thank you." - **上下文感知 PR 管理**:自动关联当前任务分支对应的 GitHub PR,实时同步状态与摘要。 - **闭环评审与自动修复**:CI 报错与 Review 建议可直接由模型原地生成修复补丁;支持行内评审及直接 Merge。 - **OpenAI Build Week**:线上线下同步联动,将在全球近 30 个城市落地举行。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=eiQgljOrkWU&t=0s) You've helped shape Codex at every step, so before I show you anything, thank you. — 主讲人 > [00:46](https://www.youtube.com/watch?v=eiQgljOrkWU&t=46s) Ultra will obviously be using your token limits faster. — 主讲人 > [01:28](https://www.youtube.com/watch?v=eiQgljOrkWU&t=88s) a smarter screenshot — 主讲人 > [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) how do I put it on the internet? — 主讲人 > [03:17](https://www.youtube.com/watch?v=eiQgljOrkWU&t=197s) publish this to Sites — 主讲人 > [06:35](https://www.youtube.com/watch?v=eiQgljOrkWU&t=395s) As always, keep giving us your hardest bugs, your boldest ideas, and the work that feels slightly impossible. Thank you. — Codex ## Mentioned OpenAI, Codex, ChatGPT, GPT-5.6 Sol, Ultra 模式, Computer Use, Browser Use, Appshot, Sites, Linear, Git, Git Worktree, Deep Research, SSH, Dev Box, GitHub, OpenAI Build Week, Passkey ## Video description Here’s what’s new in Codex for developers. Thanks for helping us make it better! We packed in a lot: * GPT‑5.6 * Ultra * New ways to manage parallel work * Big updates to Computer Use * Inline code and document edits * PR reviews * Sites * ...and much more. Make sure you watch until the end: something that isn’t out yet makes a cameo! Chapters: 00:00 What's new for developers 00:46 GPT-5.6 Sol and Ultra mode 01:28 Computer Use and appshots 02:11 Browser Use and inline editing 03:17 Publishing with Sites 03:51 Coordinating work across projects 04:57 Codex on mobile 05:36 Pull request workflows 06:35 OpenAI Build Week [Source](https://www.youtube.com/watch?v=eiQgljOrkWU)

Source quotes

[0:00] 主讲人: You've helped shape Codex at every step, so before I show you anything, thank you. [0:46] 主讲人: Ultra will obviously be using your token limits faster. [1:28] 主讲人: a smarter screenshot [3:17] 主讲人: how do I put it on the internet? [3:17] 主讲人: publish this to Sites [6:35] Codex: As always, keep giving us your hardest bugs, your boldest ideas, and the work that feels slightly impossible. Thank you.

Source quotes

[0:00] Over 6 million developers now use Codex on a weekly basis. [0:00] Codex and ChatGPT are merging into a unified application. [0:46] When combined with the Goal command, Ultra equips Sol with a significantly larger reasoning budget to tackle challenging, long-deferred projects. [0:46] Operating in Ultra mode will consume token limits at a noticeably faster rate. [1:28] An appshot functions as an intelligent screenshot, supplying Codex with both visual screen data and underlying app context. [1:28] Navigate through every screen and capture the App Store screenshots in English and French [3:17] Publish this to Sites [3:51] Find five bugs from Linear, create a separate task and isolated worktree for each one, and pin the most critical ones. [6:35] OpenAI Build Week, featuring online programming as well as nearly 30 in-person events around the world. [6:35] Your hardest bugs, your boldest ideas, and the work that feels slightly impossible.

Original transcript

[0:00] This is a big moment for Codex and I'm really [0:02] excited to show you what's new for developers. [0:04] Today, more than 6 million [0:06] of you use Codex every week. [0:08] Along the way, you've shared feedback, [0:10] shown us what you're building, [0:11] and you set an incredibly high bar. [0:14] You've asked Codex to ship ambitious projects [0:16] and help you stay organized beyond the code. [0:19] You've helped shape Codex at every step, [0:21] so before I show you anything, thank you. [0:25] Codex and ChatGPT are now coming together into one app. [0:28] For developers, it's the same Codex you know and love. [0:31] It now lives right here in ChatGPT with its own [0:34] dedicated space next to the new work agent. [0:36] We've been shipping new improvements every week, [0:38] more than 150 updates over the past [0:40] two months, and we're just getting started. [0:43] Let me show you a few of my favorites. [0:45] First, the model. [0:47] GPT-5.6 Sol is our new frontier model [0:49] and it's available to everyone. [0:51] It can work through hard problems for a very long time. [0:54] And you can also use the new [0:55] Ultra mode right here with GPT-5.6. [0:59] Combined with the Goal command, [1:00] you can give it the most challenging projects [1:02] you've been putting off, like this one. [1:04] With Ultra, Sol gets a much larger reasoning budget. [1:07] But be careful, Ultra will obviously [1:09] be using your token limits faster. [1:11] Now, as you can see here, [1:13] Codex will automatically split the work across [1:15] subagents without you needing to ask. [1:17] So I can follow any of them [1:19] individually or open this panel [1:21] and see the whole team working together. [1:23] One goal and all of this work now moving in parallel. [1:26] That's really awesome. [1:28] Sol is also outstanding at computer and browser use. [1:31] Of everything in this release, [1:32] this may be the feature I'm [1:33] the most excited for you to experience. [1:35] It's so much faster with GPT-5.6. [1:38] Say I'm working on this iPhone app, for instance. [1:40] I can simply take an appshot [1:42] of the Simulator by pressing Cmd-Cmd on Mac, [1:44] just like that, and say, navigate [1:47] through every screen and capture [1:49] the App Store screenshots in English and French. [1:52] You can think of an appshot as a smarter screenshot. [1:55] Codex gets both the screen and the app context. [1:58] With this, [2:00] Codex can use the Simulator itself. [2:01] Check this out. [2:02] This is not my cursor. [2:04] It's Codex clicking around and doing [2:05] all of this in the background where [2:07] I can continue to use my computer. [2:09] Isn't that amazing? [2:10] Switching gears, let's take a look at this very [2:13] simple paper plane game built with GPT-5.6. [2:16] I can open it directly in the in-app browser, [2:19] play with it, and see exactly what Codex created. [2:22] If I want to change something, [2:24] I can point and click here to annotate [2:26] the exact part I'm talking about. [2:28] Codex gets that visual context [2:30] to make the changes that I request. [2:32] We added a lot of features to the browser, [2:34] by the way, so you can now [2:35] work with applications that require login [2:37] as well, with support for passkeys. [2:40] And boom, here is the change. [2:42] I can also switch into the diff view. [2:44] I've always been able to inspect the diff [2:46] here, but now I can go further [2:48] and make precise edits directly inline. [2:51] So if I know exactly what I want, [2:53] I can change the line of code myself, just like that. [2:56] Or if I know the behavior I want, [2:58] but I'm not sure what the code change should be, [3:01] I can ask Codex to work from there. [3:03] Sometimes I want to work visually. [3:05] Sometimes I want to work directly into the code. [3:08] And it feels much closer [3:09] to actually pairing with the model. [3:11] Once I have something I'm happy with, [3:13] the next question is, [3:14] how do I put it on the internet? [3:16] That's where Sites comes in. [3:18] Last month, we introduced Sites for Enterprise, [3:20] so teams could create applications and share them internally. [3:23] Now, Sites is available to everyone, [3:26] and it works for any web [3:27] application that you build with Codex. [3:29] I can simply say, [3:31] publish this to Sites. [3:34] It has hosting, authentication, [3:37] a persistent database, and even file [3:39] storage, all built in. [3:40] So I can go from an idea [3:42] to a real full-stack application without really [3:44] thinking about infrastructure or hosting. [3:47] And there it is. [3:49] It's now deployed. [3:51] Zooming out a little bit, as you [3:53] delegate work to Codex across many projects, [3:55] we're continuing to make it easier [3:57] to see everything going on at once. [3:59] In fact, did you know that Codex can also [4:01] help you manage its own Codex tasks? [4:03] This is super useful when I [4:05] have several independent pieces of work [4:07] that I want to get organized and moving. [4:09] Check out this one. [4:11] Find five bugs from Linear, [4:13] create a separate task and isolated worktree for each one, [4:17] and pin the most critical ones. [4:20] And boom! [4:21] From this coordinating thread, I can see [4:24] now everything being kicked off and moving along. [4:27] So Codex is not only doing the work. [4:29] It's helping me keep all of it organized and connected. [4:32] The cool thing is, [4:33] these tasks can also reference one another. [4:36] Better yet, the work does not have to start in Codex. [4:39] If an idea starts in ChatGPT, or if I [4:42] explore it here with deep research, I [4:45] can now @mention that conversation [4:48] and bring all of the context straight [4:50] into Codex when I'm ready to build. [4:52] So no more copy-pasting. [4:53] And I'm not tied to my desk to manage Codex. [4:57] ChatGPT on mobile just got a big set of upgrades. [5:00] There are three things I wanted to try. [5:02] First, you can create, search, [5:05] open, manage tasks directly [5:07] from a conversation on your phone. [5:10] Second, reviewing changes is also much better. [5:13] I can filter between staged, [5:16] unstaged, branch, or compare [5:18] branches directly on the go. [5:20] And third, I can connect directly [5:22] to an SSH host, like a dev box, [5:25] and start new tasks from my phone. [5:27] There's so much more, but I love [5:29] that I can start something at my desk, [5:31] keep it moving from my phone, [5:33] and jump in whenever Codex needs me. [5:35] Isn't that great? [5:36] Back to desktop, we've also [5:38] made your day-to-day workflow smoother. [5:41] Codex now makes it easier to work with pull requests [5:43] without leaving the app. [5:45] It does this in context. [5:47] It knows which branch each task is working on. [5:50] So when that branch has an open PR, [5:52] the summary pane surfaces it automatically. [5:54] I can follow everything, address any issues, [5:57] and merge once everything looks good. [6:00] If a check fails, though, or a comment comes in, [6:03] I can ask Codex to fix it [6:04] right here with the full context. [6:06] The new Pull Requests tab [6:08] lets me explore every code change. [6:10] I can leave inline comments, [6:13] complete a full review. [6:14] I can also mark my own PRs as ready [6:17] and merge them. [6:18] All of this within Codex without having to go [6:21] to GitHub.com so I can stay in the flow. [6:24] This is only a small part of what we've shipped, [6:27] but the thing that matters most hasn't changed. [6:29] We're constantly improving Codex for developers, [6:32] and we're lucky to build it alongside you. [6:34] We hope you'll join us for OpenAI Build Week, [6:37] either online or in-person [6:39] at one of nearly 30 events around the world. [6:42] Okay, Codex, I'm done with the demo. [6:45] Can you get developers ready [6:46] for Build Week and wrap us up? [6:49] As always, keep giving us [6:51] your hardest bugs, your boldest ideas, [6:54] and the work that feels slightly impossible. [6:57] Thank you.