# ChatGPT and Codex (/claude-and-chatgpt-2/architecture/chatgpt-and-codex) --- type: Document title: "ChatGPT and Codex" description: "How ChatGPT, Work and Codex compare with Claude. It covers the Codex harness, a side-by-side table, and what a reported ChatGPT inspection did and did not show." status: stable order: 0.48 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-11-09T20:42:36Z chapter: "F00A" part: Front expert_status: required concepts: [ "ChatGPT and Codex" ] last_verified: 2026-11-09 sources: - id: cowork-architecture title: "Claude Cowork architecture overview (Anthropic, verified 2026-11-09)" resource: https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview - id: cloud-scheduled-tasks title: "Automate work with scheduled tasks (Anthropic, verified 2026-11-09)" resource: https://code.claude.com/docs/en/routines - id: desktop-scheduled-tasks title: "Schedule recurring tasks in Claude Code Desktop (Anthropic, verified 2026-11-09)" resource: https://code.claude.com/docs/en/desktop-scheduled-tasks - id: codex-harness title: "Unlocking the Codex harness: how we built the App Server (OpenAI, February 2026, verified 2026-11-09)" resource: https://openai.com/index/unlocking-the-codex-harness/ - id: chatgpt-agent title: "Introducing ChatGPT agent (OpenAI, July 2025, verified 2026-11-09)" resource: https://openai.com/index/introducing-chatgpt-agent/ - id: openai-sandboxing title: "ChatGPT sandbox docs (OpenAI, verified 2026-11-09)" resource: https://learn.chatgpt.com/docs/sandboxing - id: work-cloud-security title: "ChatGPT Work cloud security (OpenAI, verified 2026-11-09)" resource: https://learn.chatgpt.com/docs/enterprise/chatgpt-work-cloud-security - id: agents-api title: "Agents API architecture (OpenAI, verified 2026-11-09)" resource: https://developers.openai.com/api/docs/guides/agents-api/architecture - id: chatgpt-inspection title: "ChatGPT's inspection of its own Python and terminal environment, November 9, 2026, as reported to Panaversity" resource: "Reported to Panaversity by ChatGPT, November 9, 2026" - id: openai-tasks title: "Scheduled tasks (OpenAI, verified 2026-11-09)" resource: https://learn.chatgpt.com/docs/automations - id: whats-new title: "ChatGPT: What's new (OpenAI, verified 2026-11-09)" resource: https://learn.chatgpt.com/docs/whats-new generated: at: 2026-11-09T20:42:36Z by: process:panaversity trust_tier: unverified build_id: sha256:46fd53c5d7ab1c14f6347908d45c9fdaecda4d914439923ee878bdcecd57803e dirty: true ksor_version: 0.0.60 --- ### What is known, and how well - **[Vendor-documented]** OpenAI documents the Codex harness and its App Server. Codex Web runs that harness inside a container it provides for the task.[^codex-harness] - **[Vendor-documented]** ChatGPT Work's cloud tasks use the same harness. "Work Cloud uses the Codex task-execution harness." Work and Codex "share core execution and isolation mechanisms".[^work-cloud-security] But their tools, permissions and controls "aren't identical". - **[Unknown]** Whether regular ChatGPT chat runs on the Codex harness. Chat, Work and Codex remain separate experiences, even where their infrastructure overlaps. The Tour describes the user's side of [where ChatGPT's work runs](../tour/chatgpt.md). ### ChatGPT's four tiers | Tier | Where the harness runs | Compute environment | Persistence | Basis | | --- | --- | --- | --- | --- | | Regular ChatGPT chat | Not published | A Python sandbox (Code Interpreter), called as a tool | Not published | Not published | | ChatGPT agent[^chatgpt-agent] | Not published | "Its own virtual computer": browsers, a terminal and API access | Keeps context for the task. It can be paused and resumed | Vendor-documented | | Codex (in the ChatGPT app) | The Codex harness: a cloud container for cloud tasks, or your machine for local ones | Workspace, shell, files, MCP and skills | Threads are stored on the server | Vendor-documented[^codex-harness],[^openai-sandboxing] | | ChatGPT Work | The Codex task-execution harness, for cloud tasks | Cloud or local Work tasks | Cloud Work conversations sync across devices. Local ones stay on your computer | Vendor-documented[^work-cloud-security],[^whats-new] | ### How the Codex harness is built OpenAI's write-up describes a setup close to the Claude Code one in this section.[^codex-harness] - **The harness** is "the agent loop and logic that underlies all Codex experiences." Codex core runs the loop, saves each **thread** (one conversation) and runs tools in a sandbox. - **App Server** wraps the core in a long-running process with a two-way JSON-RPC API. It can pause a turn to ask your app for approval. - **Cloud (Codex Web).** "A worker provisions a container with the checked-out workspace, launches the App Server binary inside it." "Keeping state and progress on the server means work continues even if the tab disappears." Whether Work uses exactly this container path is not published.[^work-cloud-security] - **Local (desktop app and IDEs).** The App Server runs on your machine. Commands run in an operating-system sandbox: Seatbelt on macOS, bubblewrap on Linux and WSL2, and the native sandbox on Windows.[^openai-sandboxing] ### Side by side | Dimension | Claude (cloud session) | ChatGPT: Codex, with Work where documented | | --- | --- | --- | | Product experience | One conversation for chat and tasks, plus a Code tab, in one app. The chat examined ran on the cloud runtime behind those tasks (observed) | Chat, Work and Codex as separate experiences in one app | | Agent loop (harness) | Claude Code (observed) | Codex core in the App Server (documented for Codex and Work's cloud tasks) | | Where the computing runs | A sandbox for each session (documented)[^cowork-architecture]. A Firecracker microVM (observed) | A container for each task (documented for Codex Web). Work's cloud tasks use VM-backed sandboxes that can be reused or replaced (documented)[^work-cloud-security] | | Local option | Tasks in a conversation: local sessions only for tasks started on the device before October 6, 2026, on Pro, Max and Ultra (documented)[^cowork-architecture]. Code tab and CLI: Local, SSH and WSL, with Claude Code on your machine (documented) | Desktop app, CLI and IDE, with an operating-system sandbox (documented) | | Regular chat on the harness? | Yes, in the session examined (observed) | Unknown | | Conversation persistence | Saved to the account (documented) | Thread history on the server (documented for Codex) | | Scheduled tasks | A fresh session for each run. Cloud runs start in auto mode. Desktop runs happen on your machine (documented)[^cloud-scheduled-tasks],[^desktop-scheduled-tasks] | Run in the cloud, and can start on Gmail, Slack or GitHub events (documented)[^openai-tasks] | | Use of your computer | Through the desktop app, connected folders only (documented)[^cowork-architecture] | Work's cloud tasks use an "online, connected computer". Work Cloud cannot reach it while it is offline (documented)[^work-cloud-security] | ![Claude and Codex: same four layers, different evidence. Left, the Claude cloud session examined: the Claude app (documented), the Anthropic backend saving sessions and files (documented), the harness Claude Code in a Firecracker microVM in the session examined (observed), and the Claude model (inferred). Right, a Codex cloud task: the ChatGPT app (documented), the Codex backend storing thread history (documented), the harness Codex App Server in a container a worker provides (documented), and the OpenAI model (inferred). Notes say Work's cloud tasks use the Codex harness, and the figure compares layers, not behavior.](img/claude-and-codex-same-layers.png) Both stacks have the same shape: an app, a backend, a harness in a cloud sandbox, and the model. The difference is the evidence. For Claude, the harness was observed. For Codex, OpenAI documents it. ### A reported inspection of a ChatGPT execution environment **Source and limits.** On November 9, 2026, at our request, ChatGPT ran commands in its own Python and terminal tools, in a separate conversation. It reported the results. Nobody independently reviewed the raw output. So its findings carry the label Reported by ChatGPT.[^chatgpt-inspection] - **[Reported by ChatGPT]** The environment ran Linux x86_64 (kernel 6.18.57) with Python 3.13.5. PID 1 was `supervisord`, a program that starts and watches other programs. The root file system was an overlay with paths that referred to Kata Containers, a container technology. - **[Reported by ChatGPT]** A file written from the Python tool under `/mnt/data` could be read from a later terminal call. So working files carry over between tool calls. That does not show they survive a new conversation or a reset. - **[Reported by ChatGPT]** No Codex App Server process, and no model-serving process, was visible in its process namespace. - **[Unknown]** Where ChatGPT's conversation harness and its inference run, and when its environments are created and destroyed. A missing Codex process here does not show that Codex parts are unused elsewhere. | Question | Claude (session examined) | ChatGPT (inspection reported) | | --- | --- | --- | | Harness process visible? | Yes: Claude Code | No | | Evidence of a VM or container | Firecracker microVM (PID 1 is `/process_api --firecracker-init`) | Overlay file system with Kata Containers paths. `supervisord` as PID 1 | | Files survive a replaced VM? | Restored as copies from the conversation (observed) | Not tested | So the comparison still rests partly on documentation. **[Vendor-documented]** OpenAI's Agents API also treats the hosted harness and the execution environment as separate parts. An agent that only calls outside services "may not need its own compute or files".[^agents-api] The lesson holds for both vendors. Four jobs are separate: model inference, the agent loop, running tools, and keeping durable state. Platforms can place them behind different boundaries. [^codex-harness]: As verified on November 9, 2026. Unlocking the Codex harness: how we built the App Server, OpenAI. [^chatgpt-agent]: Introducing ChatGPT agent, OpenAI. [^openai-sandboxing]: ChatGPT sandbox docs, OpenAI. [^chatgpt-inspection]: ChatGPT's inspection of its own Python and terminal environment, as reported to Panaversity. The raw output was not independently reviewed. [^work-cloud-security]: As verified on November 9, 2026. The page says "Work Cloud uses the Codex task-execution harness", that Work and Codex "share core execution and isolation mechanisms", and that Work Cloud "manages its own execution environment". ChatGPT Work cloud security, OpenAI. [^agents-api]: As verified on November 9, 2026. The page describes the hosted harness, the environment and the application server as separate parts. Agents API architecture, OpenAI. [^cowork-architecture]: As verified on November 9, 2026. Claude Cowork architecture overview, Anthropic. [^cloud-scheduled-tasks]: As verified on November 9, 2026. Automate work with scheduled tasks, Anthropic. [^desktop-scheduled-tasks]: As verified on November 9, 2026. Schedule recurring tasks in Claude Code Desktop, Anthropic. [^openai-tasks]: As verified on November 9, 2026. Scheduled tasks, OpenAI. [^whats-new]: As verified on November 9, 2026. ChatGPT: What's new, OpenAI.