# Codex

Published January 1, 2026 · Updated August 10, 2026

AI coding assistant and IDE

Category: Developer Tools · Paid · [Official website](https://openai.com)

## Install with Homebrew

`brew install --cask codex`

## Quick take

Buy Codex when you already write ticket-quality specs and review PRs like a staff engineer. In August 2026 the stack is GPT-5.6 Sol, Terra, and Luna; you reach it through ChatGPT desktop after the 2026-07-09 merge (the Codex app for macOS/Windows named on Plus), plus web, CLI, and IDE surfaces. Sandboxes and parallel queues are the overnight backlog tool. Start on Plus; move to Pro 5× or 20× when five-hour windows and weekly caps choke daily agent farms. Skip it as your only agent if most of your week is exploratory debugging. Claude Code and Cursor Composer 2.5 still feel better for that. Keep Codex if you want a ChatGPT-billed PR factory and can live with mandatory human review. Muse Code is a Meta alternative to evaluate, not a reason to ignore GPT-5.6 routing.

Best for: Teams with well-defined routine backlogs and GitHub PR culture, Solo builders who need overnight mechanical throughput, Managers raising velocity without immediate headcount, Maintainers clearing straightforward issues at scale, Orgs already standardised on ChatGPT Plus/Pro/Business

## What is OpenAI Codex?

What is OpenAI Codex in August 2026? It is an autonomous coding agent product on the GPT-5.6 family (Sol, Terra, and Luna). You describe a software task; Codex plans, edits, runs tests in an isolated sandbox, iterates on failures, and typically opens a GitHub pull request you review like a teammate's work. It is separate from the discontinued 2021 code-completion model. Packaging also changed: the dedicated Codex macOS desktop app was merged into ChatGPT's unified desktop shell (Chat + Work + Codex) on 2026-07-09. Searching "Codex app" still means this agent product; ChatGPT's Codex pricing page lists "Build with the Codex app for macOS and Windows" under Plus, meaning the ChatGPT-hosted Codex surface, not a resurrected separate SKU. If you are choosing Codex or Claude Code: pick Codex for fire-and-forget cloud PRs on ChatGPT plans; pick Claude Code for interactive terminal steering in your real environment. Codex vs Cursor Composer 2.5: Composer wins IDE-native loops; Codex wins async sandbox queues. Codex vs Muse Code: Muse is Meta's August 2026 terminal agent; Codex remains the OpenAI-native PR factory. Model framing for search freshness: GPT-5.6 Sol, Terra, and Luna reached general availability on 2026-07-09. Sol is the hard-reasoning lane; Terra the balanced default; Luna the volume lane after OpenAI's ~80% Luna and ~20% Terra price-performance cuts around 2026-07-30. Content that still leads with GPT-5.5 as the primary Codex story is stale even if GPT-5.5 remains visible in some limit tables. Where it fits: well-specified routine engineering such as CRUD endpoints, coverage gaps, migrations, dependency bumps, and mechanical refactors. Ambiguous requirements, architecture taste calls, and repos without tests are where it stalls. Think of it as a capable junior who works overnight and still needs senior review.

## Features

- **Autonomous task execution on GPT-5.6.** Assign a scoped engineering task and let Codex plan, implement, test, and iterate without constant supervision. GPT-5.6 Sol handles deep multi-step reasoning; Terra covers balanced agent runs; Luna targets cheaper high-volume tickets after the late-July price cuts. This is delegation with a visible plan, not autocomplete.
- **Sandboxed cloud environments.** Each task gets an isolated container with your repo cloned in. Codex can install dependencies, run builds and tests, and fail safely without touching your laptop. When finished you get commits on a branch, no orphaned processes or polluted local state.
- **GitHub pull requests and code review.** Codex creates branches, writes descriptive commits, opens PRs, and can participate in review workflows (including mention-driven review patterns where enabled). You keep normal CI and human approval gates; never auto-merge agent output on protected branches.
- **Parallel multi-task queues.** Run independent tickets simultaneously in separate sandboxes. End-of-day backlog dumps, five well-written specs, often become five morning PRs. Dependent tasks should stay sequential so later work sees merged earlier work.
- **Multi-surface access: desktop, web, CLI, IDE.** After the July 2026 desktop rebuild, Codex lives in ChatGPT for Mac/Windows alongside Chat and Work, on chatgpt.com, in CLI tooling, and via IDE extensions. Start a thread in one surface and continue monitoring elsewhere depending on OpenAI's sync features for your plan.
- **Self-testing and iterative repair.** Codex does not only emit diffs. It runs your project's tests, reads failures, and patches until green or stuck. Repos with solid coverage get disproportionately better PRs because the agent can verify its own work.
- **Interactive refinement after the first PR.** Chat follow-ups, "reuse existing auth middleware," "add rate-limit errors", keep task context so refinement is incremental rather than a full restart. Useful when the first pass is 80% right.
- **Progress visibility and Work adjacency.** Watch file reads, commands, and test loops in the UI. Adjacent ChatGPT Work mode handles non-coding professional agents in the same desktop shell, reducing app sprawl for people who previously juggled separate Codex and chat clients.

## How to use OpenAI Codex in August 2026

Codex is a multi-surface product. Most Mac users should start with the official ChatGPT desktop app (unified Chat + Work + Codex). Web, CLI, and IDE paths remain for people who live in the browser or terminal. There is no current standalone Codex-only macOS app SKU.

1. **Choose a surface.** Desktop: install ChatGPT for Mac/Windows (Homebrew cask `chatgpt` or official download). Web: open Codex via ChatGPT. CLI/IDE: follow OpenAI's current Codex CLI and extension docs. Avoid obsolete "openai-codex" standalone app instructions.
2. **Sign in on a ChatGPT plan with Codex access.** Free/Go: limited trial-style Codex. Plus (~$20/mo): expanded Codex for regular use. Pro from ~$100/mo (5×) or ~$200/mo (20×): maximum task headroom. Business/Enterprise for teams. Confirm live entitlements on chatgpt.com/codex/pricing.
3. **Connect GitHub.** Authorize repository access needed for clones, branches, and PRs. Prefer least privilege, specific repos over org-wide when policy requires. Codex needs push rights to open pull requests.
4. **Pick a repository and write a sharp task.** Select a repo with README, conventions, and tests. Write acceptance criteria, constraints, file hints, and examples. Vague prompts waste credits; ticket-quality prompts save review cycles.
5. **Run, monitor, review.** Submit and watch the sandbox or walk away. When the PR lands, run CI, review diffs like a senior, refine via chat if needed, then merge. Never auto-merge solely because tests passed.

## Pros

- True async autonomy, assign, walk away, review a PR
- Cloud sandboxes isolate risk from your local machine
- Parallel task queues shrink routine backlogs overnight
- Native GitHub branch/PR workflow fits existing review culture
- GPT-5.6 Sol/Terra/Luna covers hard reasoning and cheaper volume lanes
- Multi-surface: ChatGPT desktop, web, CLI, and IDE, not a single brittle app
- Self-test loops raise quality when repos have coverage
- Unified with ChatGPT Work after July 2026, reducing OpenAI app sprawl
- Scales team throughput on predictable work without immediate hiring

## Cons

- Requires clear specs, ambiguous tasks produce thrash and burned credits
- Plan quotas (especially Plus) still constrict heavy daily agent farms; Pro is expensive
- Cloud isolation limits access to private networks, local daemons, and some secrets workflows
- Human review remains mandatory; tests do not catch every design mistake
- GitHub-centric PR story may not fit GitLab/Bitbucket-first orgs without workarounds
- Interactive exploration still often feels better in Claude Code or Cursor Composer

## How Codex works under the hood (August 2026)

Modern Codex combines GPT-5.6-class models with an agent loop tuned for software trajectories: plan, read, edit, execute, observe, repair. When you submit a task, the system analyses repository structure, READMEs, manifests, configs, representative sources, to infer stack and conventions. Visible plans let you redirect before expensive implementation begins.

Execution occurs in a sandboxed environment seeded from your connected repo. The agent installs dependencies with your project's package manager, applies edits, and frequently runs tests or typechecks. Failures feed back into the model context so patches are grounded in real stderr, not vibes. Isolation improves security and reproducibility: each run starts clean rather than inheriting a developer's half-broken local state.

Model choice shapes behaviour and cost. Sol leans into harder multi-hop reasoning and careful tool use for thorny bugs. Terra aims at balanced quality/latency for everyday agent work. Luna maximises throughput per dollar, especially salient after OpenAI's late-July price-performance push cutting Luna sharply and Terra meaningfully. Product surfaces may default differently by ChatGPT plan; power users should explicitly route volume work away from Sol when policy allows.

Delivery usually ends in Git commits and a pull request with narrative of decisions. Code review features can consume the same credit pools as implementation tasks depending on plan rules, budget review automation alongside build automation. The July 2026 ChatGPT desktop rebuild placed this loop beside Chat and Work so professionals stop alt-tabbing between a chat client and a dedicated Codex binary. CLI and IDE extensions extend the same agent identity into terminals and editors for people who never leave those habitats.

Limitations are structural. Sandboxes cannot magically reach air-gapped production. Ambiguous product goals still require humans. Benchmarks and demos overstate reliability on messy private monorepos with sparse tests. Governance, SSO, retention, training opt-outs, belongs in Business/Enterprise conversations before you connect the crown jewels. Competitors (Claude Code, Cursor Composer 2.5, Copilot agentic CLI, Muse Code) pressure OpenAI to keep shipping; evaluate quarterly, not annually.

Operationally, successful teams treat Codex as a queueing system with SLAs. They define ticket shapes that are eligible for autonomy (clear inputs/outputs, local testability, no unresolved product questions) and shapes that are banned (security-sensitive auth changes without human design, speculative refactors, anything requiring production data). They also instrument outcomes: percentage of PRs merged without refinement, median time-to-first-review, and credit burn per merged change. Those metrics catch silent quality regressions when model defaults shift after GPT-5.6 updates.

Security reviewers should map Codex to existing AI tool policies. Sandbox isolation reduces laptop malware risk but does not eliminate training, retention, or exfiltration questions for proprietary code. Business and Enterprise contracts, SSO, and retention controls matter before connecting private monorepos. Pair that with branch protection rules so agent pushes cannot skip required reviewers. On the desktop side, remember the July 2026 merge: training materials that still screenshot a separate Codex.app confuse new hires and create shadow IT installs of obsolete builds.

Finally, keep a competitive calendar. Claude Code limit changes, Cursor Composer releases, Copilot agentic CLI updates, and Meta Muse Code's August 2026 launch all change the relative value of staying OpenAI-centric. Re-evaluate quarterly with the same golden-path tickets so comparisons stay apples-to-apples on GPT-5.6 Sol/Terra/Luna versus rival stacks.

Query variants this page must answer in plain language: what is openai codex (async GPT-5.6 agent with sandboxes and PRs); codex app mac (use ChatGPT desktop Codex after the 2026-07-09 merge); codex pricing (ChatGPT plan ladder, not a separate SKU); codex vs claude code (async PR factory vs interactive terminal); codex vs cursor (ChatGPT sandbox agent vs IDE-native Composer 2.5); codex vs muse code (OpenAI vs Meta terminal/agent stacks launched into the same 2026 Mac conversation). Dealbreakers: no clear acceptance criteria, no tests, no GitHub permissions for PRs, and regulated code without Business/Enterprise review. If those are true, fix the repo and policy first. Operational metric that matters: percentage of Codex PRs merged without a refinement loop. If that number is low, write better tickets before buying Pro 20×. On Mac specifically, install or update the ChatGPT desktop app, open the Codex surface, connect GitHub with least privilege, and run one tiny task before connecting crown-jewel monorepos. Uninstall obsolete standalone Codex.app shortcuts from internal wikis so new hires do not chase a dead binary.

## FAQ

### What models does Codex use in August 2026?

Codex runs on the GPT-5.6 family, Sol, Terra, and Luna, after GA on 2026-07-09. Sol for hard reasoning, Terra balanced, Luna for cheaper volume (further improved by late-July price cuts). Do not treat GPT-5.5 or older chat stacks as the current primary story.

### Is there still a standalone Codex macOS app?

No as a separate current product. On 2026-07-09 OpenAI merged it into the ChatGPT desktop app (Chat + Work + Codex). Use ChatGPT desktop, web, CLI, or IDE extensions instead.

### How much does Codex cost?

Via ChatGPT plans: Free/Go limited trial; Plus ~$20/mo expanded Codex; Pro from ~$100/mo (5×) or ~$200/mo (20×) for maximum tasks; Business/Enterprise available. See chatgpt.com/codex/pricing and openai.com/chatgpt/pricing. Developer metering: developers.openai.com/codex/pricing.

### How is Codex different from ChatGPT chat?

Chat is conversational. Codex is an autonomous software agent that operates on repos in sandboxes, runs commands/tests, and opens PRs. Work mode sits between for broader multi-step professional tasks in the same desktop shell.

### Can Codex access my local files?

Cloud sandbox tasks work on connected repositories, not arbitrary local disks. CLI/IDE surfaces may operate closer to your machine, follow current OpenAI docs for each surface's trust boundary.

### How long do tasks take?

Simple fixes often finish in minutes; medium features in tens of minutes; larger multi-file work can approach an hour. Parallelism helps throughput more than staring at a single run.

### What if the PR is wrong?

Refine in-thread with sharper constraints, or close and rewrite the spec. Most failures trace to underspecified acceptance criteria, missing tests, or undocumented conventions.

### Codex vs Claude Code vs Cursor Composer 2.5?

Codex: async OpenAI sandboxes and PRs on GPT-5.6. Claude Code: interactive terminal with Anthropic models. Cursor Composer 2.5: IDE-native long agent runs. Pick based on whether you want fire-and-forget PRs, local steering, or editor immersion, many pros keep two.

### What about Meta Muse Code?

Muse Code launched August 2026 as a Meta terminal coding agent for macOS/Linux. It is a competitor in the agent CLI category; it does not replace Codex inside ChatGPT for teams standardised on OpenAI.

### Did Luna's price cut change how I should use Codex?

Yes for volume. After ~2026-07-30, Luna became much cheaper (~80% cut announced in OpenAI's price-performance update; Terra ~20%). Route repetitive tickets to Luna when available; reserve Sol for hard problems and watch ChatGPT plan caps either way.

### Is the old 2021 Codex model the same product?

No. The 2021 Codex completion model is historical. Modern "Codex" means OpenAI's autonomous coding agent product on GPT-5.6 surfaces.

### Does Codex replace human engineers?

No. It multiplies throughput on well-specified work and still needs review, product sense, and architecture ownership from humans.

### Codex or Claude Code: which should I use?

Use Codex when you want async sandboxes and GitHub PRs billed on ChatGPT plans with GPT-5.6. Use Claude Code when you want interactive terminal control in your real local toolchain with Anthropic models. Many teams keep both and route tickets by shape.

### Is the Codex Mac app still a separate download?

Not as a separate current product SKU. On 2026-07-09 OpenAI merged the dedicated Codex desktop app into ChatGPT desktop (Chat + Work + Codex). Plus pricing copy still mentions the Codex app for macOS and Windows as the ChatGPT-hosted surface. Install ChatGPT for Mac, not an obsolete openai-codex cask.

### Codex vs Cursor Composer 2.5 for Mac developers?

Composer 2.5 is IDE-native inside Cursor with its own Pro $20+ pools. Codex is ChatGPT-native async agents with cloud sandboxes and PR delivery. Pick Composer when you live in the editor; pick Codex when you want overnight ticket queues beside Chat and Work.

### What are Codex usage limits on Plus?

OpenAI publishes shared five-hour windows (local messages and cloud tasks) that vary by model. The Plus table on chatgpt.com/codex/pricing shows example local-message ranges such as GPT-5.6 Sol about 15-90, Terra about 20-110, and Luna about 50-280 per 5h, with weekly caps possible. Check the live table for your plan tab.

### Codex vs Muse Code?

Muse Code is Meta's August 2026 macOS/Linux terminal coding agent. Codex is OpenAI's ChatGPT-connected autonomous agent with sandbox/PR workflows on GPT-5.6. Compare them if you are vendor-shopping; they are not drop-in replacements.

## Sources

- [Codex pricing (ChatGPT)](https://chatgpt.com/codex/pricing/)
- [Introducing GPT-5.6](https://openai.com/index/gpt-5-6/)
- [Advancing the price-performance frontier with GPT-5.6](https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/)
- [What's new | ChatGPT Learn](https://learn.chatgpt.com/docs/whats-new)
- [ChatGPT macOS app release notes](https://help.openai.com/en/articles/9703738-chatgpt-macos-app-release-notes)
- [Codex developer pricing](https://developers.openai.com/codex/pricing)
- [ChatGPT pricing](https://openai.com/chatgpt/pricing)
- [Cursor Composer 2.5](https://cursor.com/blog/composer-2-5)
- [Meta Muse Code launch coverage](https://9to5mac.com/2026/08/05/meta-launches-muse-code-ai-coding-agent-for-macos-and-linux/)
- [Anthropic Claude Opus 5](https://www.anthropic.com/news/claude-opus-5)

## Related

- [Cursor](https://bundl.run/apps/cursor)
- [Claude Code](https://bundl.run/apps/claude-code)
- [ChatGPT](https://bundl.run/apps/chatgpt)
- [Claude](https://bundl.run/apps/claude)
- [Windsurf](https://bundl.run/apps/windsurf)
- [Kiro](https://bundl.run/apps/kiro)
- [Codex vs Claude Code](https://bundl.run/compare/claude-code-vs-codex)
- [Codex vs GitHub Copilot CLI](https://bundl.run/compare/codex-vs-copilot-cli)

```json
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://bundl.run/#organization",
      "name": "Bundl.run",
      "url": "https://bundl.run",
      "logo": {
        "@type": "ImageObject",
        "url": "https://bundl.run/og-image.png",
        "width": 1200,
        "height": 630
      },
      "description": "The Ninite for Mac. Install all your essential Mac apps with one terminal command.",
      "sameAs": [
        "https://github.com/abhiofficial/bundl-mac-setup",
        "https://x.com/bundlrun",
        "https://www.producthunt.com/products/bundl-run"
      ],
      "foundingDate": "2024",
      "contactPoint": {
        "@type": "ContactPoint",
        "contactType": "customer support",
        "url": "https://bundl.run/faq"
      }
    },
    {
      "@type": "WebSite",
      "@id": "https://bundl.run/#website",
      "name": "Bundl.run",
      "url": "https://bundl.run",
      "description": "The Ninite for Mac. Install all your essential Mac apps with one terminal command.",
      "publisher": {
        "@id": "https://bundl.run/#organization"
      },
      "inLanguage": "en-US"
    },
    {
      "@type": "Person",
      "@id": "https://bundl.run/authors/alex-chen#person",
      "name": "Alex Chen",
      "jobTitle": "Senior Developer Tools Specialist",
      "url": "https://bundl.run/authors/alex-chen",
      "worksFor": {
        "@id": "https://bundl.run/#organization"
      },
      "description": "Alex Chen has been evaluating developer tools and productivity software for over 12 years, with deep expertise in code editors, terminal emulators, and development environments. As a former software engineer at several Bay Area startups, Alex brings hands-on experience with the real-world workflows these tools are meant to enhance. Alex tests each application extensively on both Intel and Apple Silicon Macs, documenting performance metrics, integration capabilities, and workflow efficiency. When not reviewing software, Alex contributes to open-source projects and writes technical tutorials for the developer community.",
      "knowsAbout": [
        "Code Editors & IDEs",
        "Terminal Emulators",
        "Version Control Tools",
        "DevOps & CI/CD",
        "API Development",
        "Performance Benchmarking"
      ],
      "image": "https://bundl.run/authors/alex-chen.svg"
    },
    {
      "@type": "BreadcrumbList",
      "@id": "https://bundl.run/apps/codex#breadcrumb",
      "itemListElement": [
        {
          "@type": "ListItem",
          "position": 1,
          "name": "Home",
          "item": "https://bundl.run"
        },
        {
          "@type": "ListItem",
          "position": 2,
          "name": "Apps",
          "item": "https://bundl.run/apps"
        },
        {
          "@type": "ListItem",
          "position": 3,
          "name": "Codex",
          "item": "https://bundl.run/apps/codex"
        }
      ]
    },
    {
      "@type": "WebPage",
      "@id": "https://bundl.run/apps/codex",
      "url": "https://bundl.run/apps/codex",
      "name": "Codex for Mac",
      "description": "AI coding assistant and IDE",
      "isPartOf": {
        "@id": "https://bundl.run/#website"
      },
      "publisher": {
        "@id": "https://bundl.run/#organization"
      },
      "inLanguage": "en-US",
      "datePublished": "2026-01-01T00:00:00Z",
      "dateModified": "2026-08-10T21:30:10.000Z",
      "author": {
        "@id": "https://bundl.run/authors/alex-chen#person"
      },
      "mainEntity": {
        "@id": "https://bundl.run/apps/codex#software"
      },
      "breadcrumb": {
        "@id": "https://bundl.run/apps/codex#breadcrumb"
      }
    },
    {
      "@type": "SoftwareApplication",
      "@id": "https://bundl.run/apps/codex#software",
      "name": "Codex",
      "description": "AI coding assistant and IDE",
      "applicationCategory": "DeveloperApplication",
      "operatingSystem": "macOS",
      "url": "https://bundl.run/apps/codex",
      "image": {
        "@type": "ImageObject",
        "url": "https://img.logo.dev/openai.com",
        "caption": "Codex app icon for Mac"
      },
      "sameAs": [
        "https://openai.com/",
        "https://formulae.brew.sh/cask/codex"
      ],
      "offers": {
        "@type": "Offer",
        "price": "0",
        "priceCurrency": "USD",
        "availability": "https://schema.org/InStock"
      },
      "review": {
        "@type": "Review",
        "author": {
          "@id": "https://bundl.run/authors/alex-chen#person"
        },
        "reviewBody": "Buy Codex when you already write ticket-quality specs and review PRs like a staff engineer. In August 2026 the stack is GPT-5.6 Sol, Terra, and Luna; you reach it through ChatGPT desktop after the 2026-07-09 merge (the Codex app for macOS/Windows named on Plus), plus web, CLI, and IDE surfaces. Sandboxes and parallel queues are the overnight backlog tool. Start on Plus; move to Pro 5× or 20× when five-hour windows and weekly caps choke daily agent farms. Skip it as your only agent if most of your week is exploratory debugging. Claude Code and Cursor Composer 2.5 still feel better for that. Keep Codex if you want a ChatGPT-billed PR factory and can live with mandatory human review. Muse Code is a Meta alternative to evaluate, not a reason to ignore GPT-5.6 routing."
      },
      "about": {
        "@type": "Thing",
        "name": "Codex",
        "description": "AI coding assistant and IDE"
      },
      "isPartOf": {
        "@id": "https://bundl.run/#website"
      }
    },
    {
      "@type": "Article",
      "@id": "https://bundl.run/apps/codex#article",
      "headline": "Codex for Mac — Full Review & Installation Guide 2026",
      "description": "Buy Codex when you already write ticket-quality specs and review PRs like a staff engineer. In August 2026 the stack is GPT-5.6 Sol, Terra, and Luna; you reach it through ChatGPT desktop after the 2026-07-09 merge (the Codex app for macOS/Windows named on Plus), plus web, CLI, and IDE surfaces. Sandboxes and parallel queues are the overnight backlog tool. Start on Plus; move to Pro 5× or 20× when five-hour windows and weekly caps choke daily agent farms. Skip it as your only agent if most of your week is exploratory debugging. Claude Code and Cursor Composer 2.5 still feel better for that. Keep Codex if you want a ChatGPT-billed PR factory and can live with mandatory human review. Muse Code is a Meta alternative to evaluate, not a reason to ignore GPT-5.6 routing.",
      "image": "https://img.logo.dev/openai.com",
      "author": {
        "@id": "https://bundl.run/authors/alex-chen#person"
      },
      "publisher": {
        "@id": "https://bundl.run/#organization"
      },
      "datePublished": "2026-01-01T00:00:00Z",
      "dateModified": "2026-08-10T21:30:10.000Z",
      "mainEntityOfPage": {
        "@type": "WebPage",
        "@id": "https://bundl.run/apps/codex"
      },
      "articleSection": "DeveloperApplication",
      "speakable": {
        "@type": "SpeakableSpecification",
        "cssSelector": [
          "h1",
          ".key-facts"
        ]
      },
      "about": {
        "@type": "SoftwareApplication",
        "name": "Codex",
        "url": "https://bundl.run/apps/codex"
      }
    },
    {
      "@type": "FAQPage",
      "@id": "https://bundl.run/apps/codex#faq",
      "mainEntity": [
        {
          "@type": "Question",
          "name": "What models does Codex use in August 2026?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Codex runs on the GPT-5.6 family, Sol, Terra, and Luna, after GA on 2026-07-09. Sol for hard reasoning, Terra balanced, Luna for cheaper volume (further improved by late-July price cuts). Do not treat GPT-5.5 or older chat stacks as the current primary story."
          }
        },
        {
          "@type": "Question",
          "name": "Is there still a standalone Codex macOS app?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "No as a separate current product. On 2026-07-09 OpenAI merged it into the ChatGPT desktop app (Chat + Work + Codex). Use ChatGPT desktop, web, CLI, or IDE extensions instead."
          }
        },
        {
          "@type": "Question",
          "name": "How much does Codex cost?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Via ChatGPT plans: Free/Go limited trial; Plus ~$20/mo expanded Codex; Pro from ~$100/mo (5×) or ~$200/mo (20×) for maximum tasks; Business/Enterprise available. See chatgpt.com/codex/pricing and openai.com/chatgpt/pricing. Developer metering: developers.openai.com/codex/pricing."
          }
        },
        {
          "@type": "Question",
          "name": "How is Codex different from ChatGPT chat?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Chat is conversational. Codex is an autonomous software agent that operates on repos in sandboxes, runs commands/tests, and opens PRs. Work mode sits between for broader multi-step professional tasks in the same desktop shell."
          }
        },
        {
          "@type": "Question",
          "name": "Can Codex access my local files?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Cloud sandbox tasks work on connected repositories, not arbitrary local disks. CLI/IDE surfaces may operate closer to your machine, follow current OpenAI docs for each surface's trust boundary."
          }
        },
        {
          "@type": "Question",
          "name": "How long do tasks take?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Simple fixes often finish in minutes; medium features in tens of minutes; larger multi-file work can approach an hour. Parallelism helps throughput more than staring at a single run."
          }
        },
        {
          "@type": "Question",
          "name": "What if the PR is wrong?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Refine in-thread with sharper constraints, or close and rewrite the spec. Most failures trace to underspecified acceptance criteria, missing tests, or undocumented conventions."
          }
        },
        {
          "@type": "Question",
          "name": "Codex vs Claude Code vs Cursor Composer 2.5?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Codex: async OpenAI sandboxes and PRs on GPT-5.6. Claude Code: interactive terminal with Anthropic models. Cursor Composer 2.5: IDE-native long agent runs. Pick based on whether you want fire-and-forget PRs, local steering, or editor immersion, many pros keep two."
          }
        },
        {
          "@type": "Question",
          "name": "What about Meta Muse Code?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Muse Code launched August 2026 as a Meta terminal coding agent for macOS/Linux. It is a competitor in the agent CLI category; it does not replace Codex inside ChatGPT for teams standardised on OpenAI."
          }
        },
        {
          "@type": "Question",
          "name": "Did Luna's price cut change how I should use Codex?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Yes for volume. After ~2026-07-30, Luna became much cheaper (~80% cut announced in OpenAI's price-performance update; Terra ~20%). Route repetitive tickets to Luna when available; reserve Sol for hard problems and watch ChatGPT plan caps either way."
          }
        },
        {
          "@type": "Question",
          "name": "Is the old 2021 Codex model the same product?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "No. The 2021 Codex completion model is historical. Modern \"Codex\" means OpenAI's autonomous coding agent product on GPT-5.6 surfaces."
          }
        },
        {
          "@type": "Question",
          "name": "Does Codex replace human engineers?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "No. It multiplies throughput on well-specified work and still needs review, product sense, and architecture ownership from humans."
          }
        },
        {
          "@type": "Question",
          "name": "Codex or Claude Code: which should I use?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Use Codex when you want async sandboxes and GitHub PRs billed on ChatGPT plans with GPT-5.6. Use Claude Code when you want interactive terminal control in your real local toolchain with Anthropic models. Many teams keep both and route tickets by shape."
          }
        },
        {
          "@type": "Question",
          "name": "Is the Codex Mac app still a separate download?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Not as a separate current product SKU. On 2026-07-09 OpenAI merged the dedicated Codex desktop app into ChatGPT desktop (Chat + Work + Codex). Plus pricing copy still mentions the Codex app for macOS and Windows as the ChatGPT-hosted surface. Install ChatGPT for Mac, not an obsolete openai-codex cask."
          }
        },
        {
          "@type": "Question",
          "name": "Codex vs Cursor Composer 2.5 for Mac developers?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Composer 2.5 is IDE-native inside Cursor with its own Pro $20+ pools. Codex is ChatGPT-native async agents with cloud sandboxes and PR delivery. Pick Composer when you live in the editor; pick Codex when you want overnight ticket queues beside Chat and Work."
          }
        },
        {
          "@type": "Question",
          "name": "What are Codex usage limits on Plus?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "OpenAI publishes shared five-hour windows (local messages and cloud tasks) that vary by model. The Plus table on chatgpt.com/codex/pricing shows example local-message ranges such as GPT-5.6 Sol about 15-90, Terra about 20-110, and Luna about 50-280 per 5h, with weekly caps possible. Check the live table for your plan tab."
          }
        },
        {
          "@type": "Question",
          "name": "Codex vs Muse Code?",
          "acceptedAnswer": {
            "@type": "Answer",
            "text": "Muse Code is Meta's August 2026 macOS/Linux terminal coding agent. Codex is OpenAI's ChatGPT-connected autonomous agent with sandbox/PR workflows on GPT-5.6. Compare them if you are vendor-shopping; they are not drop-in replacements."
          }
        }
      ],
      "isPartOf": {
        "@id": "https://bundl.run/apps/codex"
      }
    },
    {
      "@type": "ItemList",
      "name": "Codex Video Tutorials",
      "description": "Tutorial videos for learning Codex on Mac",
      "itemListElement": [
        {
          "@type": "ListItem",
          "position": 1,
          "item": {
            "@type": "VideoObject",
            "name": "OpenAI Codex Tutorial #1 - Introduction & Setup",
            "description": "In this Codex tutorial series, you'll learn how to use the Codex Cloud platform to spin up remote tasks, the Codex CLI to work locally on projects, and the Codex IDE Extension to use Codex directly wi",
            "thumbnailUrl": "https://i.ytimg.com/vi/tIb_TzVNbDM/mqdefault.jpg",
            "uploadDate": "2025-09-29T07:20:34Z",
            "duration": "PT9M29S",
            "embedUrl": "https://www.youtube.com/embed/tIb_TzVNbDM",
            "interactionStatistic": {
              "@type": "InteractionCounter",
              "interactionType": "https://schema.org/WatchAction",
              "userInteractionCount": 92452
            }
          }
        },
        {
          "@type": "ListItem",
          "position": 2,
          "item": {
            "@type": "VideoObject",
            "name": "GPT-5 Codex: From Beginner to Expert in 17 minutes",
            "description": "In this video I go over an INCREDIBLE workflow for GPT-5 Codex! Timestamps: 0:00 Intro 1:11 Setup 3:47 Building an app 9:09 Integrating results 14:32 Marketing agents",
            "thumbnailUrl": "https://i.ytimg.com/vi/e-poAzH8Zck/mqdefault.jpg",
            "uploadDate": "2025-09-18T20:50:51Z",
            "duration": "PT16M57S",
            "embedUrl": "https://www.youtube.com/embed/e-poAzH8Zck",
            "interactionStatistic": {
              "@type": "InteractionCounter",
              "interactionType": "https://schema.org/WatchAction",
              "userInteractionCount": 77628
            }
          }
        },
        {
          "@type": "ListItem",
          "position": 3,
          "item": {
            "@type": "VideoObject",
            "name": "OpenAI Codex in your code editor",
            "description": "Use Codex right in your code editor with the new IDE extension for VS Code, Cursor and other forks. In this video, Gabriel Peal and Romain Huet show you how to explore code, implement new features, an",
            "thumbnailUrl": "https://i.ytimg.com/vi/sd21Igx4HtA/mqdefault.jpg",
            "uploadDate": "2025-10-17T14:01:34Z",
            "duration": "PT7M1S",
            "embedUrl": "https://www.youtube.com/embed/sd21Igx4HtA",
            "interactionStatistic": {
              "@type": "InteractionCounter",
              "interactionType": "https://schema.org/WatchAction",
              "userInteractionCount": 199390
            }
          }
        },
        {
          "@type": "ListItem",
          "position": 4,
          "item": {
            "@type": "VideoObject",
            "name": "Getting started with Codex",
            "description": "Get started with Codex, OpenAI’s coding agent, in this step-by-step onboarding walkthrough. You’ll learn how to install Codex, set up the CLI and VS Code extension, configure your workflow, and use Ag",
            "thumbnailUrl": "https://i.ytimg.com/vi/px7XlbYgk7I/mqdefault.jpg",
            "uploadDate": "2026-01-12T16:11:51Z",
            "duration": "PT53M2S",
            "embedUrl": "https://www.youtube.com/embed/px7XlbYgk7I",
            "interactionStatistic": {
              "@type": "InteractionCounter",
              "interactionType": "https://schema.org/WatchAction",
              "userInteractionCount": 108553
            }
          }
        },
        {
          "@type": "ListItem",
          "position": 5,
          "item": {
            "@type": "VideoObject",
            "name": "How To Use Codex CLI: Complete Beginner Guide & Tips",
            "description": "In this tutorial I teach you how to use Codex CLI from OpenAI to it's full potential. AI coding CLI's like Claude Code and Codex have quickly become essential coding tools so in this video I give you ",
            "thumbnailUrl": "https://i.ytimg.com/vi/sTE0G95uEIw/mqdefault.jpg",
            "uploadDate": "2025-11-24T18:00:42Z",
            "duration": "PT1H3M42S",
            "embedUrl": "https://www.youtube.com/embed/sTE0G95uEIw",
            "interactionStatistic": {
              "@type": "InteractionCounter",
              "interactionType": "https://schema.org/WatchAction",
              "userInteractionCount": 3728
            }
          }
        }
      ],
      "numberOfItems": 5
    }
  ]
}
```