Codex
FreemiumCodex started life in April 2025 as OpenAI's coding agent: a command-line tool and, later, a cloud-based agent that could take a plain-language instruction and turn it into a pull request, working independently while a developer went and did something else. A little over a year later, the product has grown into something considerably broader. OpenAI now describes Codex simply as your AI assistant for work, and the shift is not just marketing copy: internally at OpenAI, the median employee across engineering, legal, finance, and recruiting now generates the majority of their output through Codex rather than through ChatGPT itself.\n\nWhat changed is the scope of tasks Codex is trusted with. Where the original release was squarely about writing and debugging code, Codex in 2026 can be handed research questions, spreadsheets, briefs, prototypes, and recurring operational work, then asked to gather the source material, do the work, and hand back something reviewable, a document, a deck, a fix, an automation, rather than just a diff. It runs across three connected surfaces: the standalone Codex app for multi-tasking across projects, an editor extension, and a terminal-based CLI, all backed by the same account and session history, so a task started in one place can be picked up in another.\n\nThe part that most separates Codex from a typical coding assistant is how it handles long, unsupervised work. Tasks run in isolated worktrees so multiple agents can work on the same repository without stepping on each other, and OpenAI's own data shows the type of request has shifted meaningfully over the past year: by mid-2026 the majority of individual user requests correspond to tasks that would take a human more than 30 minutes, with a fast-growing share representing multi-hour or even multi-day work. Automations let Codex run on a schedule in the background, handling things like daily issue triage or CI-failure summaries, with finished output landing in a review queue rather than requiring live supervision. Skills, bundles of instructions, scripts, and resources, extend it further into non-coding territory: OpenAI has demonstrated Codex building a browser game end to end using an image-generation skill and a web-game-development skill, working independently across millions of tokens.\n\nFor developers specifically, Codex still functions as a serious coding agent in the tradition of GitHub Copilot or Cursor, but leans further toward autonomous execution than either: reviewing pull requests, scanning for vulnerabilities, and running workflows locally, in the cloud, or against tools like Slack and Linear rather than staying purely inline in an editor. The non-developer adoption curve is the more striking data point, though. OpenAI reports non-developer individual usage growing over 130 times since mid-2025, with legal and recruiting staff now using Codex to automate spreadsheet work, structured analysis, and internal tooling that would previously have gone to an engineering team.\n\nCompared to standalone coding-agent products like Cursor or Devin, Codex's built-in advantage is distribution: anyone already using ChatGPT can turn it on without a separate signup, separate billing relationship, or separate account to manage, and the same session history and skills carry across the CLI, the desktop app, and mobile. That convenience comes with a tradeoff, since Codex's roadmap and pricing move in lockstep with the broader ChatGPT product rather than shipping as an independent release cycle, so features and limits can shift with little warning whenever OpenAI updates a ChatGPT plan.\n\nThe honest limitation is less about capability and more about trust and oversight. Codex's own adoption data inside OpenAI shows a genuine shift in how engineering teams work, code review increasingly happens agent-to-agent, and the human bottleneck moved from writing code to QA capacity, and that pattern will not suit every team's risk tolerance, particularly for anyone who needs a human in the loop on every change by default.\n\nAccess follows the ChatGPT plan you are already on: the Free tier includes limited Codex access, Plus (20 dollars a month) covers focused weekly coding sessions, Pro (starting at 100 dollars a month) gives higher usage for full workdays across multiple projects, and Business moves to pay-as-you-go per-seat pricing with shared workspaces and admin controls, with Enterprise offering custom deployment, advanced analytics, and broader integration support. For a solo developer or a small team already paying for ChatGPT Plus, Codex is effectively already included and worth turning on; for teams weighing whether to adopt it as their primary way of working, OpenAI's own internal numbers are as good a case study as any of what that transition looks like in practice.