OpenAI Codex

What it is

OpenAI Codex is OpenAI’s coding agent. You describe a task, and it reads the repository, edits files, runs commands and tests, and hands back a diff or a pull request. It runs on six surfaces: the ChatGPT desktop app (macOS, Windows and, since August 2026, Linux), ChatGPT on the web, the ChatGPT mobile app, the open-source Codex CLI, an IDE extension, and Codex Cloud, which works through tasks in OpenAI-hosted environments while your laptop is closed. On 9 July 2026 OpenAI folded Codex into the redesigned ChatGPT desktop app next to ChatGPT Work, its agent for non-coding tasks, and the two now draw on one usage pool.

Adoption moved fast. OpenAI reported more than 3 million weekly users in April 2026 and more than 5 million in June; by late July it was quoting about 10 million weekly users for Codex and ChatGPT Work combined, without splitting the two.

This page covers the coding agent. ChatGPT covers the assistant itself.

Why it shows up in ops stacks

Most ops teams already pay for ChatGPT. When the company is on ChatGPT Business or Enterprise, the GTM engineer or RevOps developer gets Codex in the same seat, with no new vendor review. Four capabilities make it useful beyond that:

• Cloud tasks and GitHub review. Codex Cloud runs a task against a connected GitHub repository without a local checkout, and GitHub code review — generally available — can review pull requests automatically or when someone comments @codex review. GitLab support is in beta.
• Work starts where the request lands. The Slack and Linear integrations let someone hand Codex a task from a thread or an issue, which suits the “can someone fix the lead-routing script” request that arrives in a channel.
• Headless and embeddable. Non-interactive mode (codex exec), a GitHub Action and the Codex SDK put the same agent into CI and cron jobs. MCP support connects it to CRM, ticketing and warehouse servers; see what an MCP server is.
• AGENTS.md. Codex reads repository instructions from AGENTS.md, the shared convention that Claude Code also falls back to, so one file can brief both agents.

Pricing reality

Codex has no list price of its own. It comes with ChatGPT plans:

• Free and Go ($8/mo) — a light model in the desktop app only, subject to rollout. Not a working tool.
• Plus — $20/mo. Web, CLI, IDE extension, iOS and cloud integrations.
• Pro — $100, $200 or $500 per month, with no five-hour limit.
• Business — $25 per user per month, or $20 billed annually, minimum two users; adds admin controls and SAML SSO.
• Enterprise and Edu — custom pricing, with SCIM, RBAC, the Compliance API and data residency.
• API key — pay API rates for the CLI, SDK and IDE extension, without cloud features such as GitHub review or Slack.

Since 2 April 2026, usage beyond a plan’s allowance is billed in credits per million tokens, not per message. OpenAI’s pricing page says a typical task on its mid-tier model uses 2–15 credits; third-party pricing guides put a purchased credit at about $0.04, so roughly 8–60 cents a task. Plus and standard Business seats carry five-hour and weekly caps, and the page’s own estimates span a wide range per model.

One path has closed. In April OpenAI let Business workspaces add Codex-only, pay-as-you-go seats; on 24 June 2026 it stopped selling new ones, though existing seats continue. A team that wants Codex without ChatGPT for everyone now buys standard Business seats or runs on an API key with a project budget. Plan on $20–25 per light user per month, and $100–200 per month for anyone who codes with it all day.

Best for

GTM engineers, RevOps developers and legal-ops or recruiting-ops technologists at companies that have standardized on ChatGPT Business or Enterprise. The scoped case where it wins: delegating well-defined tasks — a failing sync job, a field-mapping change, a new webhook handler — from a Slack thread or Linear issue to a cloud task, then reviewing the pull request it opens. The Business seat the ops team already holds covers that case with no extra procurement.

Not for

People who cannot judge a diff, and teams without a git workflow. Codex hands back code changes; if nobody reviews them, give that person ChatGPT Work or a no-code builder instead. It is also the wrong home for high-volume, fixed-shape jobs: once the job is “sync these fields every hour,” the code Codex wrote should run in n8n or your orchestrator, not in agent sessions billed by the token.

Versus the alternatives

• GitHub Copilot — the largest installed base by seat count. Business is $19 and Enterprise $39 per user per month, with metered usage above the included credits since 1 June 2026. Pick it when the organization already buys GitHub Enterprise and the main need is inline completion inside the editor for many developers.
• Cursor — the top coding-specific rival by revenue, at $20/mo for individuals and $40 per user for Teams. Pick it when the person lives in an IDE and wants to steer each edit visually; see alternatives to Cursor.
• Claude Code — Anthropic’s agent, in Claude plans from $20/mo, with $100–200 Max and $125 Team premium seats for heavy users. Pick it when the job is long multi-file changes and scheduled runs, or the company is on Claude rather than ChatGPT. Codex is itself the fastest-growing entrant in this segment, so the deciding question is usually which assistant the company already pays for.

The full field is ranked in best AI coding assistants.

Watch-outs

• One pool, many consumers. Codex shares its allowance with ChatGPT Work and voice tasks, so a colleague’s research report draws on the same limit as your refactor. Guard: run CI and cron jobs on an API key with its own spend limit, and on credit-based Business workspaces set workspace credit limits before turning Work on for everyone.
• Models retire on a schedule you do not set. GPT-5.5 leaves Codex on all plans on 14 October 2026, and older Codex model snapshots were shut down in the API on 23 July. Pinned model names in a shared config.toml or a CI job stop working. Guard: avoid pinning snapshot names in shared configuration, check OpenAI’s deprecations page monthly, and re-run a small set of known tasks after each model change.
• Approval modes go further than they look. “Approve for me” routes eligible actions to automatic review instead of a person, and “Full access” removes the sandbox altogether. Guard: start in “Ask for approval,” as OpenAI’s own docs recommend, use managed requirements on Business and Enterprise to block Full access for repositories that touch production, and require human approval on any MCP tool that writes to the CRM.