OpenAI’s agentic coding product. A coding agent that works either async in the cloud (fire a task, come back to a PR) or locally via the Codex Desktop app, CLI, and IDE extensions. Backed by GPT-5.5 for harder work and GPT-5.3-Codex-Spark for faster routine coding.
Bundled with every paid ChatGPT tier. Pro tiers get 5x to 20x the Codex usage of Plus. The April 16, 2026 super-app update turned Codex Desktop into the most capable OpenAI product surface to date: Computer Use, persistent Memory, gpt-image-2 for visuals, an in-app browser, 90+ plugins, and multi-agent workflows running in parallel.
Recent developments (March-May 2026)
- May 2, 2026: Codex Desktop added Pets, Hatch custom pet generation, cross-agent config imports, and a dictation dictionary. The playful overlay is less important than the config portability: Codex can now import conventions from tools such as Claude Code instead of forcing users to rewrite setup context.
- May 1, 2026: The MCP STDIO command-execution flaw put agent connectors back under security review. Codex Desktop’s plugin/MCP layer should be treated as privileged tool access, not as harmless extension metadata.
- May 1, 2026: The Pentagon expanded classified-network AI access to eight major vendors, including OpenAI. This is a government deployment story rather than a public Codex feature, but it reinforces why account security, network boundaries, and audit controls matter for agentic OpenAI products.
- April 30, 2026: OpenAI added Advanced Account Security for ChatGPT and Codex. Hardware-key and passkey protection now matter more because Codex accounts can touch repos, terminals, plugins, and connected development workflows.
- April 30, 2026: OpenAI began rolling GPT-5.5 Cyber to critical defenders through Trusted Access. The relevant Codex angle is controlled access to cyber-permissive capabilities, not broad public availability.
- April 30, 2026: A coding-agent security roundup warned that attackers keep targeting credentials, not model weights. Codex evaluations should include sandbox isolation, token scopes, account security, and audit trails alongside model quality.
- April 30, 2026: OpenAI and Microsoft gut exclusivity, freeing OpenAI to serve products across AWS and Google Cloud. Codex now has a clearer path to non-Azure enterprise distribution, but practical availability still depends on future product rollouts.
- April 30, 2026: OpenAI’s DevDay 2026 is set for September 29. Codex roadmap and product direction will likely be a centerpiece of the event.
- April 29, 2026: Agent skill libraries are becoming the new coding-agent workflow layer. Codex benefits from the same shift toward reusable local skills, templates, and MCP-aware workflow modules instead of one-off prompts.
- April 28, 2026: Codex comes to Amazon Bedrock as part of the expanded OpenAI/AWS partnership. Limited preview starts with Codex CLI, the desktop app, and the VS Code extension configured to use Bedrock as the provider.
- April 27, 2026: OpenAI and Microsoft made their model partnership non-exclusive. Codex remains part of the OpenAI product stack, but enterprise deployment optionality around OpenAI products now matters more than Azure exclusivity alone.
- April 27, 2026: Musk v. OpenAI trial opens with fraud claims dismissed. No immediate Codex feature changes, but governance disclosures and remedies could still affect OpenAI’s operating model.
- April 27, 2026: OpenAI publishes new principles for the next AGI phase. Developers should watch how those principles translate into model access, safety gates, enterprise controls, and Codex release notes.
- April 25, 2026: AI News Desk, April 25 grouped GPT-5.5 API access, Copilot distribution, Project Deal, and Google-Anthropic financing as the weekend signals around Codex’s market.
- April 24, 2026: AI Industry Roundup, April 24 tracked GPT-5.5’s spread into Copilot; GitHub Copilot adds GPT-5.5 is the direct competing-coding-surface rollout.
- April 23, 2026: GPT-5.5 rolls out to ChatGPT and Codex. OpenAI positions it for longer-running coding, research, data analysis, documents, spreadsheets, and tool-use tasks.
- April 23, 2026: AI Industry Roundup, April 23 covered GPT-5.5, Cloud Next, Grok Voice, Grok outages, and EU AI funding; GPT-5.5 system card and bio bounty added safety context for Codex’s backing model.
- April 22, 2026: Responses API adds WebSockets. Relevant to Codex-style agent loops because lower-latency bidirectional sessions make browser/IDE agent workflows feel less stop-start.
- April 23, 2026: Google discloses 75% of internal new code is AI-generated. Signals agentic-coding adoption inside hyperscalers has crossed the “production default” threshold. Codex sits in the same category as Gemini Code Assist for this cohort.
- April 23, 2026: Gemini Enterprise Agent Platform launches. Google now ships a full enterprise agent stack (Studio, Registry, Identity, Gateway, Observability) competing with Codex for governed-agent use cases. Codex counter is the April 16 super-app update’s plugin ecosystem and Computer Use.
- April 21, 2026: Moonshot Kimi K2.6 ships with Agent Swarm mode, posting SWE-bench Pro 58.6 as the strongest open-weights baseline. Raises the ceiling for self-hosted alternatives to Codex.
- April 16, 2026: Codex Desktop ships as OpenAI “super app”. Computer Use, Memory, gpt-image-2, in-app browser, 90+ plugins, multi-agent workflows. macOS first; Windows and EU/UK rolling out.
- April 16, 2026: Systemic MCP vulnerability exposes 200k servers. Codex Desktop’s 90+ plugins + MCP tool access all inherit the exposure. Audit third-party servers; prefer first-party.
- April 16, 2026: Agents SDK ships native sandbox execution, model-native harness, configurable memory, snapshotting. Python first, TypeScript coming.
- April 9, 2026: $100/mo Pro tier launched. Between ChatGPT Plus ($20) and the former $200 Pro (now Pro 20x). 5x Codex vs Plus; 10x promo through May 31.
- April 2, 2026: API-aligned billing. Switched from per-message to per-token credit consumption (credits per MTok input, cached input, output).
System Verdict
Pick Codex if you want an async coding agent that runs in the background while you do other work. Codex’s “fire a task, come back to a pull request” model is the strongest pattern in the category for routine features, test generation, boilerplate, and documentation. The April 16 super-app update makes Codex Desktop the most capable all-in-one developer surface from OpenAI. Best for users already paying for ChatGPT Plus or higher.
Skip it if you need a Linux desktop GUI today, primary in-IDE coding, or the tightest terminal agent. Cursor owns daily-driver IDE coding with real-time autocomplete and in-editor chat. Claude Code owns terminal-native autonomous runs and the 1M context window for codebase-scale work. Devin owns fully-autonomous enterprise agent-first workflows. Codex is strongest as a hybrid: async cloud agent + capable desktop companion.
Who pays which tier: Plus $20/mo for most individuals doing occasional Codex tasks, Pro $100/mo for engineers running Codex weekly (5x usage, matches Claude Max 5x pricing), Pro 20x $200/mo for sustained agentic coding workloads. Business $25/user for teams. Codex-only seats are available with pay-as-you-go token billing.
Key Facts
| Default model | GPT-5.5 for harder work; GPT-5.3-Codex-Spark for faster routine coding on supported surfaces |
| Fast model | GPT-5.3-Codex-Spark (research preview, faster daily-coding) |
| Primary surfaces | Codex Desktop (macOS, Windows rolling) · ChatGPT web · Codex CLI · IDE extensions |
| Async cloud execution | Yes. Sandboxed VM spins up per task; Codex returns a diff or PR |
| Computer Use | Yes (new April 16, macOS first). Codex operates desktop apps with a virtual cursor |
| Memory | Persistent across sessions (new April 16). Remembers preferences, corrections, setup context |
| Image generation | gpt-image-2 bundled in Codex Desktop (new April 16) |
| In-app browser | Yes (new April 16). Research, form-fill, scrape without leaving the app |
| Plugins | 90+ (new April 16). Combines skills, app integrations, MCP servers |
| Multi-agent | Yes. Run multiple Codex agents concurrently across different tasks or modules |
| Billing | Credits per MTok input / cached input / output (since April 2, 2026) |
| Platform support | macOS (primary) · Windows rolling out · EU/UK rolling out · Linux via CLI |
Every data point verified 2026-04-18 against OpenAI’s Codex docs and multi-source coverage.
What it actually is
Three related surfaces united by the same underlying agent:
1. Async cloud agent (the original Codex). From inside ChatGPT, you describe a task (“add a new settings page to this repo, tests included”). Codex spins up a sandboxed VM with your codebase cloned in, works through the task independently over minutes or hours, and returns a diff or pull request. You continue with other work meanwhile.
2. Codex Desktop (the April 16 super-app). A standalone macOS app (Windows rolling out) that runs multiple Codex agents in parallel, uses Computer Use to operate applications directly, holds persistent Memory across sessions, generates images with gpt-image-2, browses the web in-app, and orchestrates 90+ plugins including first-party app integrations and third-party MCP servers.
3. Codex CLI + IDE extensions. Local-first developers use Codex as a terminal CLI (similar shape to Claude Code) or via IDE extensions in VS Code, JetBrains, and others. IDE extensions hit the same GPT-5.5 and Codex-Spark model family that backs the cloud and desktop surfaces.
The real moat is the cross-surface consistency. A task started in the desktop app can hand off to the cloud agent. Memory carries across sessions and surfaces. Plugins work everywhere. OpenAI is treating Codex as the front end of a developer-focused super app rather than a single product.
When to pick OpenAI Codex
- You already pay for ChatGPT Plus, Pro, or Business. Codex is included. Claude Code needs a separate Claude subscription; Cursor needs a separate Cursor Pro.
- You want async background work. Fire tasks, keep coding, come back to PRs. This is Codex’s strongest pattern.
- You run multiple parallel tasks. Codex’s multi-agent workflows parallelize across different modules, different repos, or different kinds of work.
- You need Computer Use. Codex Desktop operates apps directly with a virtual cursor. Useful when an MCP server or API doesn’t exist for your target app.
- You want plugin-ecosystem depth. 90+ plugins covering Figma, Linear, GitHub, Notion, Slack, Jira, and more. Fewer setup hops than assembling your own MCP stack.
- You like Memory across sessions. Cross-session context survives restarts so you don’t re-explain your setup.
When to pick something else
- Daily-driver IDE coding: Cursor. Tighter in-editor autocomplete, Composer for multi-file edits, and Cursor 3’s Agents Window for parallel agent orchestration. Works on Linux, macOS, Windows.
- Terminal-native autonomous runs: Claude Code. Strongest CLI coding agent; 1M context on Opus 4.7; Ultraplan cloud environments; no desktop-app requirement.
- Fully-autonomous enterprise agent: Devin by Cognition AI. $500/mo Teams. Agent-first from the ground up, runs end-to-end without human review.
- Open-source terminal agent: Aider. Free, self-hosted, works with any model including local Ollama.
- IDE plugin inside existing VS Code setup: GitHub Copilot. Deepest Microsoft ecosystem integration.
- Hosted full-app builder: Replit Agent. Runs the full app lifecycle on Replit infrastructure.
Pricing
Codex is bundled with every paid ChatGPT tier. No standalone Codex subscription. Usage scales with tier.
| ChatGPT tier | Monthly | Codex usage vs Plus |
|---|---|---|
| Free | $0 | Limited Codex (rate-limited) |
| Go | $8 | Basic Codex |
| Plus | $20/mo | 1x (baseline) |
| Pro | $100/mo | 5x (10x promo through May 31, 2026) |
| Pro 20x | $200/mo | 20x |
| Business | $25/user/mo | Per-seat Codex + admin controls |
| Enterprise | Custom | Custom usage + SSO + compliance |
API-aligned billing (since April 2, 2026): Credits are the purchase unit. Consumption is measured in input tokens, not per-message. Teams on pay-as-you-go get transparent cost-per-task tracking.
Prices verified 2026-04-18 via OpenAI Codex pricing and the Codex rate card.
Against the alternatives
| OpenAI Codex | Claude Code | Cursor | Devin | |
|---|---|---|---|---|
| Primary surface | Desktop app + cloud + CLI | Terminal CLI | VS Code fork IDE | Web app |
| Async cloud execution | Yes (sandboxed VM) | Ultraplan cloud environments | Cloud sandboxes + parallel agents | Yes, cloud-native by default |
| Backing model(s) | GPT-5.5 + GPT-5.3-Codex-Spark | Claude Opus 4.7 + Sonnet 4.6 | User-selectable (Claude, GPT, Gemini, Grok, Composer 2) | Proprietary Cognition models |
| Computer Use | Yes (April 16, macOS first) | Research preview in Claude Code | No native computer use | Yes, agent-native |
| Memory | Yes (April 16, persistent cross-session) | Projects + session scope | Via .cursorrules + rules | Agent-native memory |
| Image generation | gpt-image-2 bundled | None native | None native | None native |
| Plugins / ecosystem | 90+ (April 16) | MCP registry + Anthropic Skills | Extensions + custom commands | Built-in toolset |
| Platform support | macOS now; Windows + EU/UK rolling; Linux via CLI | macOS, Windows, Linux | macOS, Windows, Linux | Web (cross-platform) |
| Pricing model | Bundled with ChatGPT Plus+ | Bundled with Claude Pro+ | $20/mo Cursor Pro + API | $500/mo Teams |
| Best for | Async background work + OpenAI ecosystem | Long-context terminal agent | Primary daily-driver IDE | Fully autonomous tasks |
Failure modes
- macOS-first desktop app. Windows + EU/UK rolling out but not yet universal. Linux developers use the CLI or wait.
- Async tasks can fail silently. Cloud-sandboxed VMs sometimes error without a clear message. Check task status rather than assuming success.
- 90+ plugin surface = larger attack surface. The April 16 MCP vulnerability and the May 1 MCP STDIO command-execution disclosure affect every Codex MCP server. Audit third-party plugins and sandbox STDIO tools before trusting them with shell access.
- Pricing complexity. Free / Go / Plus / Pro / Pro 20x / Business / Enterprise is a lot of tiers. Users often pay for the wrong one. Baseline recommendation: Plus for individuals, Pro for active coders, Pro 20x only for sustained daily use.
- Codex quality lags Claude Code on strict SWE-bench. Claude Opus 4.7 leads SWE-bench Verified; Codex is close behind at ~80%. Codex wins on Terminal-Bench 2.0 at 77.3% for autonomous terminal ops.
- Computer Use is early. Works for simple app operations; fragile for complex UI automation. Test your workflow before relying on it.
- Memory is not end-to-end encrypted by default. Enterprise tiers add stronger guarantees; consumer tiers store persistent context on OpenAI’s infrastructure subject to their data retention policies.
Methodology
This page was produced by the aipedia.wiki editorial pipeline, an automated system that ingests vendor documentation and multi-source news coverage, verifies facts against primary sources, and generates the editorial analysis you are reading. No individual human wrote this review. Scoring follows the four-dimension rubric at /about/scoring/ (Utility, Value, Moat, Longevity, unweighted average). Last verified 2026-04-27 against OpenAI’s Codex docs, the Codex rate card, and multi-source coverage including the April 16 super-app update announcement.
FAQ
Is Codex a separate subscription? No. Codex is bundled with every paid ChatGPT tier: Plus ($20/mo), Pro ($100/mo), Pro 20x ($200/mo), Business ($25/user/mo), and Enterprise.
What’s the difference between Codex, ChatGPT, and Codex Desktop? ChatGPT is the general-purpose assistant. Codex is the coding agent inside ChatGPT and in standalone surfaces. Codex Desktop is the macOS app that runs Codex agents + Computer Use + 90+ plugins. Same underlying product family.
How does Codex compare to Claude Code? Both are agentic coding products. Codex works async in the cloud and via desktop app; Claude Code runs terminal-native on your local machine with the 1M Opus 4.7 context window. Claude Code is tighter for long-context refactors; Codex is stronger for async background work and plugin-driven workflows.
How does Codex compare to Cursor? Cursor is an IDE (VS Code fork) optimized for daily in-editor coding with real-time autocomplete, Composer, and the Agents Window for parallel agents. Codex is primarily async cloud + desktop app. Many developers use both: Cursor as primary IDE, Codex for fire-and-forget tasks.
Can I run Codex on Linux? Codex Desktop is macOS-first with Windows rolling out. Linux users access Codex via the Codex CLI or ChatGPT web. A native Linux desktop app has not been announced.
What’s GPT-5.3-Codex-Spark? A faster Codex-optimized model in research preview as of April 2026. Designed for day-to-day coding with lower latency than full GPT-5.5 reasoning runs. Available on Pro and Pro 20x tiers.
Does Codex support MCP? Yes. Codex Desktop’s plugin system uses MCP under the hood for tool access. Be aware of the April 16 systemic MCP vulnerability before installing third-party MCP servers.
What is Memory? Persistent context across Codex sessions. Remembers your preferences, corrections, and project setup so you don’t re-explain environment details each time. Launched April 16, 2026 on the ChatGPT-signed-in Codex Desktop.
Can Codex operate other apps? Yes via Computer Use (April 16, macOS first). Codex sees the screen, controls a virtual cursor, clicks, and types. Useful when MCP plugins or APIs don’t exist for your target app.
Is there a team tier? Yes. ChatGPT Business at $25/user/mo. Plus Codex-only seats with pay-as-you-go token billing for teams that want Codex without full ChatGPT per-user costs.
Sources
- OpenAI: Codex product page
- OpenAI: Codex pricing
- OpenAI: Codex rate card
- OpenAI: Codex flexible pricing for teams
- Codex Desktop Pets and config-import update
- Codex Desktop super-app coverage
- Builder.io: Codex vs Claude Code comparison
- The New Stack: Cursor + Claude Code + Codex merging into one AI coding stack
- DataCamp: Codex vs Claude Code
- Apidog: Cursor vs OpenAI Codex
Related
- Category: AI Coding
- Parent product: ChatGPT (Codex ships inside every paid ChatGPT tier)
- Direct competitors: Claude Code · Cursor · Devin · Aider · GitHub Copilot · Replit Agent · Windsurf
- News: Codex Desktop super-app · Pro tier launch
Embed this score on your site Free. Links back.
<a href="https://aipedia.wiki/tools/codex/" target="_blank" rel="noopener"><img src="https://aipedia.wiki/badges/codex.svg" alt="OpenAI Codex on aipedia.wiki" width="260" height="72" /></a> [](https://aipedia.wiki/tools/codex/) Badge value auto-updates if the editorial score changes. Attribution via the link is required.
Cite this page For journalists, researchers, and bloggers
According to aipedia.wiki Editorial at aipedia.wiki (https://aipedia.wiki/tools/codex/) aipedia.wiki Editorial. (2026). OpenAI Codex — Editorial Review. aipedia.wiki. Retrieved May 8, 2026, from https://aipedia.wiki/tools/codex/ aipedia.wiki Editorial. "OpenAI Codex — Editorial Review." aipedia.wiki, 2026, https://aipedia.wiki/tools/codex/. Accessed May 8, 2026. aipedia.wiki Editorial. 2026. "OpenAI Codex — Editorial Review." aipedia.wiki. https://aipedia.wiki/tools/codex/. @misc{openai-codex-editorial-review-2026,
author = {{aipedia.wiki Editorial}},
title = {OpenAI Codex — Editorial Review},
year = {2026},
publisher = {aipedia.wiki},
url = {https://aipedia.wiki/tools/codex/},
note = {Accessed: 2026-05-08}
} Spotted an error or want to share your experience with OpenAI Codex?
Every tool page is re-verified on a recurring cycle, and corrections land faster when readers flag them directly. If you spot a stale fact, a missing capability, or have used OpenAI Codex and want to share what worked or didn't, the editorial desk reviews every message sent through this form.
Email editorial@aipedia.wiki