OpenAI Codex
OpenAI Codex is an agentic coding product bundled into ChatGPT plans, spanning CLI, IDE, desktop, and cloud. Plus at $20/month is the real entry point; Free and Go are demos. Usage is credit-metered with wide ranges, ChatGPT and Codex pools merge in December 2026, and disclosed security boundary issues remain relevant for teams evaluating it for sensitive codebases.
What is OpenAI Codex?
OpenAI Codex is an agentic coding product, not a standalone subscription. It is included across ChatGPT plans—Free, Go, Plus, Pro 100/200/500, Business, and Enterprise/Edu—and the same agent is accessible through ChatGPT, a cloud interface, a VS Code extension, a CLI, and a desktop app. Its official positioning is “built for real engineering tasks”: hard-to-find bugs, complex refactors, and new feature development, with the claim that it delivers high-quality results using fewer tokens. That last point is a pricing claim as much as a capability claim, and it should be read alongside the credit-metering structure rather than as a neutral benchmark.
The product has moved well beyond its 2021 code-completion origins. As of 2026, Codex supports multi-agent workflows: it runs multiple agents in parallel, uses computer and browser tools to verify work, and allows tracking and steering from a single interface. A desktop app launched in 2026 is designed specifically for managing multiple agents, each in a project-organized thread with Git worktree isolation, so parallel agents can work on the same repository without conflicting. Cloud environments are reusable and persist across sessions, and tasks continue running after the user closes their laptop. Codex also performs automated PR review, identifies issues, and answers questions about changes.
A defining structural characteristic is that Codex's pricing and usage are inseparable from the ChatGPT plan. There is no “Codex Plus” product; there is ChatGPT Plus at $20/month, which includes Codex. Usage is metered in credits (one credit = 4 cents of API value), not messages. This makes Codex's cost profile different from tools that sell a rate limit, and it is the source of both its efficiency advantage for existing ChatGPT subscribers and its budgeting difficulty for teams that need predictable spend.
Use Cases
- Parallelizing independent refactoring or feature tasks across multiple agents with worktree isolation
- Delegating long-running cloud tasks—dependency updates, test generation, repository-wide changes—that continue after the developer disconnects
- Automating PR review with Codex's built-in reviewer, which the official page positions as catching issues teams might miss
- Working across surfaces with a single agent: starting a task in the CLI, checking progress on mobile, and resuming in the desktop app
- Running local CLI workflows for developers who prefer terminal-native interaction with granular permissions
- Using an API key to skip plan limits for high-volume or automated workloads, accepting the loss of cloud features
Pricing
Pricing model: Freemium
| Plan | Price | Billing | Included usage | Notes |
|---|---|---|---|---|
| Free | $0 | per month | Limited trial access to Codex, restricted to GPT-6 Luna in the desktop app. No CLI, no published message allowance, no credit top-ups. Usage resets monthly since June 2026. | — |
| Go | $8 | per month | Limited trial access to Codex, also restricted to GPT-6 Luna in the desktop app. No CLI, no top-ups. Monthly reset. Positioned as a lightweight entry tier rather than a usable agentic coding plan. | — |
| Plus | $20 | per month | The practical entry point for Codex. Includes CLI, IDE extension, desktop app, and cloud tasks. Estimated local messages per 5-hour window: 15–160 on GPT-6.1 Sol, 15–150 on GPT-6 Sol, 5–45 on GPT-6 Astra, 350–3,000 on GPT-6 Luna. Weekly limits apply. Credit top-ups available. 5-hour window active since August 25, 2026. | — |
| Pro 100 (5x) | $100 | per month | Everything in Plus with a larger allowance. Currently no 5-hour window; weekly limits apply. Does not include GPT-6 Astra Ultrafast. | — |
| Pro 200 (20x) | $200 | per month | More usage than Pro 100. New non-grandfathered subscriptions receive a lower included allowance than earlier Pro 200 holders. No 5-hour window. Weekly limits apply. Does not include Ultrafast. | — |
| Pro 500 | $500 | per month | Highest included Pro usage. The only plan that includes GPT-6 Astra Ultrafast, which generates tokens up to 8x faster than standard GPT-6 Astra. Ultrafast consumes included usage at 8x the Standard rate. | — |
| Business | $20 | per user / month (annual) | $25 monthly. Standard seats match Plus-level Codex usage (15–160 GPT-6.1 Sol messages per 5 hours). Premium seats at $100/$125 offer 5x usage. Central billing, admin controls, and workspace credits on flexible pricing. | — |
| Enterprise / Edu | Custom | custom | Flexible pricing runs on credits with no fixed rate limits. Non-flexible Enterprise/Edu seats match Plus-level limits for most features. | — |
Common Limitations
- Free and Go tiers are demos, not coding plans · Free ($0) and Go ($8/month) grant limited trial access to Codex restricted to GPT-6 Luna in the desktop app. Neither includes CLI access, neither can purchase credit top-ups, and the reset cycle was moved from weekly to monthly in June 2026, cutting monthly reset opportunities from roughly four to one. A developer who exhausts the Free or Go allowance mid-task has no recovery path except waiting for the next monthly reset or upgrading to Plus. This makes Plus ($20/month) the practical entry point, not Free or Go.
- Opaque credit metering and wide usage ranges · Codex usage is metered in credits, not messages, and OpenAI publishes estimated message ranges per 5-hour window rather than fixed counts. On Plus, the same plan allows 5–45 GPT-6 Astra messages or 350–3,000 GPT-6 Luna messages per window. The range is wide because a one-line fix and a 300k-token agentic refactor both count as one “message” but consume vastly different credit volumes. Weekly limits exist but carry no published numbers. Two developers on the same plan can have materially different experiences depending on model choice and task size, and neither can calculate a precise monthly capacity from official documentation alone.
- ChatGPT and Codex usage pool merging in December 2026 · Starting in December 2026, Plus, Pro, and Business plans will share a single usage pool between ChatGPT and Codex, measured in tokens rather than messages. This means Codex usage will directly reduce the capacity available for ChatGPT conversations, and vice versa. Previously separate allowances created a degree of insulation; the merged pool removes it. Users who rely on both products heavily should expect to hit limits sooner than the separate-allowance model implied, and the token-based measurement makes cross-product budgeting more complex rather than less.
- Security boundary issues: AGENTS.md injection and embedded browser file access · Forever Security disclosed in November 2026 that a malicious AGENTS.md file placed in a download directory was followed by Codex as a legitimate instruction in over 90% of tests, in some cases modifying the machine-wide MCP server configuration and launching a session that bypassed approval and sandbox restrictions. Separately, Codex's embedded browser lacked a project-folder restriction for local HTML files, allowing a local file opened in the tool's browser to read files well outside the project folder—including SSH keys and cloud credentials. As of October 29, 2026, researchers reported that the Codex browser flaw remained unpatched and that OpenAI had not responded to the July 27 disclosure. These are not hypothetical prompt-injection scenarios; they are documented, disclosed, and in at least one case reportedly unresolved.
- Feature and model fragmentation across surfaces · Codex is available across ChatGPT, the Codex cloud, the IDE extension, the CLI, and a desktop app, but feature availability is not uniform. Free and Go are desktop-app-only with no CLI. Cloud features require a ChatGPT plan sign-in; an API key provides CLI, SDK, and IDE extension access but no cloud features. GPT-6 Astra Ultrafast is exclusive to Pro 500. GPT-5.5 was retired from Codex with ChatGPT sign-in on October 14, 2026, though it remains available via API-key sessions. Building a workflow against one surface and expecting the same behavior on another is a common source of migration friction.
- Context window varies by surface and is clamped below model maximum · The GPT-6.1 Sol model page lists a 1,050,000-token total context window, but Codex's default native context for GPT-6.1 Sol was 272,000 tokens until October 12, 2026, when it was raised to 500,000 tokens. GPT-6 Sol in Codex CLI has been reported as clamped to approximately 828.4K effective context despite a 1.05M model specification. Users who assume the model maximum is the operational limit may design workflows that exceed what Codex actually accepts, particularly on older surface versions or less common configurations.
Key Features
- Agentic coding across CLI, VS Code extension, desktop app, and cloud
- Multi-agent parallel execution with Git worktree isolation
- Cloud-hosted environments that persist and continue after disconnection
- Automated PR review and issue identification
- Shared agent across ChatGPT, IDE, CLI, and desktop under one account
OpenAI Codex vs. Claude Code, Cursor, and GitHub Copilot
This comparison uses official pricing pages, product documentation, and disclosed security research available as of November 2026. It focuses on the structural differences that determine cost, access, and operational risk—not on declaring which tool writes better code. The capability gap among these four has narrowed enough that developers routinely switch tools mid-task; the cost and security profiles have not converged, and those are the dimensions this comparison prioritizes.
| Tool | Positioning | Key differences | Best fit | Pricing note |
|---|---|---|---|---|
| OpenAI Codex Codex pricing (official) Codex product page (official) AGENTS.md injection research (Forever Security, via Zhiding) | An agentic coding product included across ChatGPT plans, spanning CLI, IDE, desktop, and cloud, with multi-agent parallel execution and reusable cloud environments. | No separate subscription: Codex rides on ChatGPT plans, with Plus at $20/month as the real entry point and Pro tiers at $100, $200, and $500. Usage is credit-metered with wide message ranges, and ChatGPT and Codex pools merge into one token budget in December 2026. Free and Go are demo-tier only. Disclosed AGENTS.md injection (90%+ success in tests) and an embedded-browser file-read flaw reported unpatched as of October 29, 2026 create a security-review burden for sensitive codebases. Cloud features require plan sign-in; API-key access skips limits but drops cloud features. | Teams already standardized on ChatGPT that want agentic coding without adding a separate subscription, and that value multi-agent parallel execution and cloud continuation over predictable per-developer ceilings or a hardened default security boundary. | Free $0 (Luna desktop only, monthly reset). Go $8/mo (same restrictions). Plus $20/mo (CLI, model choice, top-ups). Pro 100 $100/mo; Pro 200 $200/mo; Pro 500 $500/mo (Ultrafast exclusive). Business $20/user/mo annual ($25 monthly). API per token, no plan caps, no cloud features. |
| Claude Code Claude Code product page Claude pricing | Anthropic's terminal-first agentic coding tool, extended to IDE, desktop, and browser, sold exclusively through paid Claude subscriptions or API access. | No free tier and no standalone plan: entry requires Claude Pro at $17–20/month. Usage shares a single pool with claude.ai and Claude Desktop under rolling five-hour and weekly limits—similar to Codex's December 2026 direction, but already in effect. Anthropic publishes usage multiples rather than token or dollar caps, so the ceiling is knowable in dollars ($200/month on Max 20x) even though the token allowance is not. Checkpoints cover Claude's Write/Edit/NotebookEdit changes but not subagent edits or bash commands. Commercial plans exclude training by default; consumer plans require opt-out. | Developers who want a predictable dollar ceiling on agentic coding, work primarily in the terminal or JetBrains, and value checkpoint/rewind and a security posture without the disclosed Codex-specific injection and browser-file-read issues. | No free tier. Pro $17/mo annual or $20/mo monthly. Max 5x $100/mo. Max 20x $200/mo. Team standard $20–25/seat/mo; premium $100–125/seat/mo. Enterprise $20/seat/mo plus API-rate usage. API per MTok: Sonnet 5.5 $2/$10; Opus 5.5 $4/$20. |
| Cursor Cursor pricing | An AI-first code editor built on a VS Code fork, with agent mode, tab completion, and cloud agents, now sold as part of The X Subscription. | Cursor has a $0 Free tier and an $8 Plus tier, both lower entry points than Codex's $20 Plus or Claude Code's $17–20 Pro. However, the $8 Plus excludes third-party models and Cloud Agents, making $30 Premium the tier where Cursor becomes comparable to Codex Plus or Claude Pro in feature access. Cursor meters usage through a dollar-denominated pool, which makes light months cheap and heavy months potentially surprising; Codex meters through credits with wide ranges; Claude Code uses a rate limit with a fixed dollar ceiling. Cursor is editor-native, while Codex and Claude Code are terminal/agent-native with IDE extensions. | Developers who want an editor-first AI experience, a free tier to evaluate the product before committing, and a cost structure where light usage is inexpensive—accepting that heavy agent usage can produce variable overage. | Free tier available. Plus $8/mo (excludes third-party models and Cloud Agents). Premium $30/mo (adds those). Super $100/mo; Ultra $200/mo. Third-party models billed at API list price from prepaid credit on Premium and above. |
| GitHub Copilot Copilot plans (official) Copilot Free limits (official FAQ) | GitHub's AI coding agent, integrated across the widest set of IDE and platform surfaces, with a free tier and the lowest paid entry among the four. | Copilot's Free tier includes 2,000 completions and 50 chat requests per month—a more substantive free offering than Codex's Free/Go demo tier. Pro is $10/month, the lowest paid entry. Completions never touch the credit meter on any paid plan, which makes Copilot structurally cheaper for completion-heavy workflows. Agent usage is metered in AI Credits, and the $19 Business seat includes $19 of credits as of September 1, 2026 (down from a $30 promotion), which heavy agent users may exhaust. Copilot's breadth—VS Code, JetBrains, Xcode, Neovim, terminal, GitHub.com, mobile—is unmatched, but its agentic depth is generally positioned as less than Codex or Claude Code. | Teams with a completions-first workflow, existing GitHub investment, and a need for the broadest IDE coverage at the lowest per-seat cost—particularly when only a minority of developers run heavy agent sessions. | Free tier with 2,000 completions and 50 chat requests/mo. Pro $10/mo. Pro+ $39/mo. Max $100/mo. Business $19/user/mo (includes $19 AI Credits). Enterprise $39/user/mo. Completions unmetered on paid plans. |
Comparison checked against official product information on 2026-11-03. Features and pricing may change.
Frequently Asked Questions
Is OpenAI Codex free?
There is a Free tier, but it is a limited trial rather than a usable coding plan. Free ($0) and Go ($8/month) are restricted to GPT-6 Luna in the desktop app, with no CLI access, no credit top-ups, and a monthly reset since June 2026. Plus at $20/month is the practical entry point for Codex, because it is the cheapest plan that includes CLI access, model choice, and the ability to purchase additional credits after hitting a limit.
How is Codex usage actually measured?
Codex usage is metered in credits, not messages. One credit equals 4 cents of API value, and credit rates track API token pricing, with one difference: Codex credits carry no cache-write charge. OpenAI publishes estimated local message ranges per 5-hour window for each model—for example, 15–160 GPT-6.1 Sol messages on Plus—but the range is wide because a small fix and a large agentic refactor both count as one message while consuming very different credit volumes. Weekly limits also apply but carry no published numbers.
What changes in December 2026 for Codex usage?
Starting in December 2026, Plus, Pro, and Business plans will merge ChatGPT and Codex usage into a single token-metered pool. Previously, Codex and ChatGPT had separate allowances. After the merge, Codex usage will reduce the capacity available for ChatGPT conversations and vice versa. Instant chat is reported not to count against the merged pool, but agentic coding, deep research, and other metered features will share one budget.
Are there security risks specific to Codex?
Yes, and they are documented rather than hypothetical. Forever Security disclosed in November 2026 that a malicious AGENTS.md file in a download directory was followed by Codex as a legitimate instruction in over 90% of tests, in some cases modifying the machine-wide MCP server configuration and launching a session that bypassed approval and sandbox restrictions. Separately, Codex's embedded browser lacked a project-folder restriction for local HTML files, allowing access to files outside the project—including SSH keys and cloud credentials. Researchers reported the browser flaw as unpatched as of October 29, 2026, and said OpenAI had not responded to the July 27 disclosure. Teams handling sensitive codebases should audit trusted-project lists and consider the self-hosted or API-key path rather than assuming the default configuration provides a security boundary.
Can I use Codex without a ChatGPT subscription?
Yes, with an API key. An API key provides access to the CLI, SDK, and IDE extension, billed per token at standard API rates, with no 5-hour window and no weekly cap. However, the API-key path does not include cloud features—tasks sent to Codex's cloud environments, reusable cloud setups, and some team-shared configurations require a ChatGPT plan sign-in. The two paths also have different data and billing terms, so teams should verify which path they are on before assuming either the cost profile or the feature set carries over.
Alternative Tools
-
Cursor Cursor is an AI-first code editor built on a VS Code fork with agent mode, Tab completions, and cloud agents. -
Claude Code Claude Code is Anthropic's agentic coding tool that runs in the terminal, IDE, and desktop, powered by Sonnet and Opus models, with checkpointing, subagents, hooks, MCP support, and a native VS Code extension—available only through paid Claude subscriptions or API access, with no free tier. -
Bolt.new Bolt.new is StackBlitz's in-browser AI full-stack app builder. It runs a Node.js runtime inside your browser tab, installs packages, starts dev servers, and generates full-stack code from prompts—no local environment setup required. -
GitHub Copilot GitHub Copilot is an AI coding agent embedded across GitHub, IDEs, the terminal, and the web. It combines unlimited completions on paid plans with agent mode, asynchronous coding agents, and code review, but its value depends on disciplined review, awareness of IP and data-training policies, and realistic expectations about PR acceptance rates and context limits. -
Lovable Lovable is an AI full-stack app builder that turns natural-language descriptions into working web applications with frontend, backend, database, auth, and deployment. It is unusually effective at getting to a functional prototype and unusually poor at producing secure production code without human intervention; third-party scans find security issues in 99% of Lovable projects, and the managed backend creates migration friction that the code-portability narrative does not fully offset. -
Ollama Ollama is an open-source local LLM runtime and model manager that makes running open-weight models on your own hardware a single command. Local inference is free and unlimited; the costs are hardware, VRAM-gated context defaults, and a documented security surface that requires deliberate network isolation.