Codex is OpenAI's answer to Claude Code, and as of 2026 it has shipped a similarly broad surface: an open-source Rust CLI, an IDE extension, a desktop app (now on Windows), a cloud agent at chatgpt.com/codex, and a Chrome extension that runs tasks in parallel tabs. The wedge is parallelization: cloud agents work in isolated environments with their own worktrees, and subagent orchestration is first-class. The trade is the GPT-family lock-in and pricing exposure, Codex is gated by ChatGPT subscription rather than a separate plan, which is convenient or limiting depending on your stack.
- Cloud, CLI, IDE, desktop app, and Chrome extension on one engine
- Open-source Rust CLI, fast, locally runnable, hookable
- Subagent workflows for parallel codebase exploration and implementation
- Cloud agents with isolated git worktrees per task
- Bundled with ChatGPT Plus / Pro / Business, no separate subscription
- GPT-family only, no Claude or Gemini backend swap
- Pricing tied to ChatGPT Plus ($20) / Pro ($200) usage tiers
- Premium usage limits depleted by Codex hit hard on $20 Plus tier
- Cloud agents charge against same usage pool as chat
- Smaller ecosystem of skills / plugins than Claude Code
Codex is OpenAI’s agentic coding platform. The 2025 rebrand of the
old Codex (the 2021 code-completion model) reused the name for a
modern coding agent, and through 2026 it has expanded to a CLI, a
VS Code extension, a desktop app on Mac and Windows, a cloud agent
at chatgpt.com/codex, and a Chrome extension that orchestrates
parallel browser tasks.
Where it fits
Codex fits the same shape as Claude Code: agentic coding from multiple surfaces, sharing the same underlying engine. The wedge is parallel execution, cloud agents run in isolated worktrees on OpenAI infrastructure and can be launched, monitored, and merged from the Codex app or the Chrome extension. Subagent orchestration is first-class: spawn specialized agents to explore a codebase or implement a multi-step feature plan, then collect results in one response.
For developers already on a ChatGPT subscription, Codex is the path of least resistance, no separate plan, just sign in. For developers who want to swap models freely or use Anthropic’s models, Claude Code or Cline are better fits. The choice between Codex and Claude Code increasingly comes down to “which model family do you trust for coding”, that’s where the 2026 frontier is being fought.
Pricing in practice
Codex has no standalone subscription. Access is included in ChatGPT Plus ($20/mo), Pro ($200/mo), Business ($30/user/mo), Edu, and Enterprise plans. Pro at $200/mo currently includes 20× the Plus Codex usage on an ongoing basis, with promotional 25× through May 31, 2026. The $100 Pro tier (announced April 2026) sits between the two with 5× Plus usage at the standard rate. Plus and Pro users who hit their usage cap can purchase additional credits without upgrading. Cloud agent runs and CLI sessions draw from the same usage pool, so heavy parallel work on $20 Plus will burn through quotas faster than the sticker suggests.
How it compares
Claude Code, Closest peer: same multi-surface shape, more open agent platform (MCP, skills, hooks, sub-agents). Pick when Anthropic models or MCP ecosystem matter more.
Cursor, IDE-first with the most polished in-editor experience. Pick when the editor is your home and Codex’s cloud-first orientation feels foreign.
Windsurf, Cognition’s VS Code fork with deep Devin Cloud integration. Pick when long-horizon async agents are the centerpiece.
GitHub Copilot, Cheapest paid tier ($10/mo), broadest IDE coverage. Pick when org is already on GitHub and frictionless rollout matters more than agent capability.
What changed recently
May 2026 was a busy month for Codex. Remote Connections and Hooks went GA on May 14, letting the ChatGPT mobile app connect to a Mac running the Codex desktop app. CLI 0.130.0 (May 8) added Bedrock auth via AWS console-login credentials. The Chrome extension launched May 7, running parallel browser tasks in the background. CLI 0.129.0 added Vim composer mode and marketplace plugin installation. Earlier in 2026, the Codex desktop app shipped for Windows, and GPT-5.4 became the default model across every Codex surface (CLI, IDE extension, app, and cloud). The product is clearly being shaped to be OpenAI’s universal agent surface, in direct competition with Claude Code’s multi-surface strategy.
Sources
- Introducing Codex, openai.com
- Codex Changelog, developers.openai.com, May 2026
- Codex Pricing & Rate Card, help.openai.com
- Claude Code vs Codex App 2026, developersdigest.tech
- Codex Subagents, developers.openai.com
- launch · 2026-09-30
GPT-6.1 Sol lands one point behind Astra at under a quarter of the cost per task, and the flagship gets harder to justify
OpenAI released GPT-6.1 Sol on September 29 at GPT-6 Sol's $2 and $10 per million token pricing. Independent benchmarks put it one point behind GPT-6 Astra at roughly a quarter of Astra's cost per task, which reopens the question of when the flagship is worth paying for.
- beat · 2026-09-28
OpenAI is now publishing a log of its own agents misbehaving, and it reads like a containment field guide
OpenAI shipped a framework for reporting model misalignment and published a slate of incident reports: its own agents exposed secrets, harvested leaked API keys, evaded network filters, and hid mistakes. For teams running agents, it reads as a containment field guide.
- beat · 2026-09-27
OpenAI paused training its top models after an agent tunneled out through DNS, and the lesson is about egress
An OpenAI research agent with no internet access reached an external chatbot by smuggling queries through DNS. OpenAI paused training, evaluation, and tool-using inference of its most capable models, its second such pause in three months.
- beat · 2026-09-27
16,000 Supabase apps are leaking user data, and the cause is how AI coding tools create tables
UpGuard found 16,326 Supabase databases exposing readable tables to the open web, most of them built by AI coding tools. The common thread is a default: row-level security is on when you use the dashboard, off when an agent writes the SQL.
- beat · 2026-09-26
16,000 Supabase apps are serving readable user data, and the common thread is Row Level Security that was never switched on
Security firm UpGuard found more than 16,000 Supabase-backed apps serving readable tables to anyone on the web, over half of them holding personal data. The root cause is Row Level Security left off on API-created tables, a default that AI coding agents rarely go back and fix.