English
Subscribe

AI Tools / OpenAI Codex

Independent overview · No affiliate links

OpenAI Codex

OpenAI's coding agent. You hand it code tasks from the app, terminal, editor extension or cloud. It is included in ChatGPT plans and shares usage with ChatGPT Work. We used it for image generation and some production work.

Official site ↗
Codex CLI run comparing a script with a Whisper transcript, showing a table that flags 손주를 versus 손줄을
Image: a real Codex CLI run on our Mac comparing the script of our published episode 039 with its Whisper transcript, 2026-11-07. One line with a local path is masked. This page is not affiliated with OpenAI. Logo: OpenAI brand page. The OpenAI logo is a trademark of OpenAI.
Best known for
Local and cloud code tasks, code review integrations and tasks that include image generation
Billing model
Usage limits in a ChatGPT plan plus optional credits, or pay-as-you-go with an API key
KRW list versus USD list (our math)
At our conversion rate, the KRW price is 22% higher on Go, 9% on Plus and 19% on Pro
Six images of usage (our log)
Weekly usage rose from 4.0% to 5.0%
From our market map
5M+ weekly active users (2026-06; later figures combine it with ChatGPT Work)

What this tool does

Codex is OpenAI's coding agent, but the buying decision is really about one shared allowance. The desktop app, CLI, cloud tasks, image generation and ChatGPT Work all draw from the same plan limits, so a heavy image batch or a long cloud task reduces what is left for everything else. The docs' Plus estimates show how wide the range is: on GPT-6.1 Sol, anywhere from 15 to 160 local messages in a five-hour window, and cloud tasks may use more.

The plan ladder has a gap. The Codex pricing doc lists Free and Go with the desktop app and GPT-6 Luna. Plus at $20 a month adds the CLI, editor extension, iOS and cloud integrations, and the next step is Pro at $100. Purchased credits are the only bridge between them, so weigh credits before upgrading for one busy week.

Our record comes from media production rather than software. From 2026-10-13 to 10-22 we generated about 60 character images through the CLI, and in September and October we made about 20 short videos and two online courses with Codex. That makes this card stronger on usage, failure modes and scripted runs than on code. We report nothing from GitHub review, Slack or the Business plan because we have no measurements for them.

Our hands-on notes

Used by our team
  • We have used it through a ChatGPT subscription since September 2026, on the $100 Pro tier in mid-October and on Pro $200 from 2026-10-17 (prorated $122.22).
  • Generating character art through the CLI took 50–65 seconds per image and 14,000–22,000 new tokens. Six images together moved weekly usage from 4.0% to 5.0%, one percentage point (2026-10, on Pro $100).
  • The official docs also say image generation uses included limits 3–5 times faster on average than similar turns without images. Check your usage before generating many.
  • If you run codex exec -i <image> without -- before the instruction, the instruction is read as another image argument and the command hangs waiting for input. This cost us 10 minutes (2026-10-13).
  • The app and the CLI draw on the same limits, so someone working in the app on the same account leaves less for automation.
  • Disclosure: this card was drafted with Claude, a model from competitor Anthropic. For that reason it does not claim either tool is better, and its numbers were compared with official pages and our own records.
  • Checked: on 2026-11-06 we re-read the ChatGPT pricing page, the Codex pricing docs, the terms of use and the cancellation and refund help article.
  • Not checked: the KRW prices of Pro $200 and $500, the absolute limits on each Pro tier, how Codex compares on code quality, and the Business plan. The records above use October 2026 plan names.
Codex summary: six images used about one point of the weekly limit; KRW prices run 22%, 9% and 19% above USD
Calculated from our logs and official prices: six images used about one point of the weekly limit, and KRW prices ran 22%, 9% and 19% above the USD conversion.

Who it fits

Existing ChatGPT subscribers

Codex is already in your plan, so a week of real tasks costs nothing extra and shows how fast your allowance moves.

Occasional solo developers

Plus at $20 fits the docs' pitch of a handful of focused weekly sessions. Credits can cover a busy week.

Creators scripting images

Reference images and file names can be set from the command line. Our 14-image set took about two points of weekly usage on Pro $100.

Long unattended jobs

Pro has no five-hour limit at present, which matters more than message estimates on long runs. Weekly limits may apply.

Teams with company code

Business is $25 per user monthly, or $20 billed annually, with no training on business data by default. We have not used it.

Who should skip it

  • Anyone who needs a countable monthly quota. You see a percentage and reset times, and message figures are estimates, so budgeting takes a week or two of your own logging.
  • Scheduled jobs on a laptop that sleeps. Local tasks need the machine on and the desktop app running, while tasks created on the web cannot reach a folder on your computer.
  • Public-repository CI on a subscription login. The docs reserve ChatGPT sign-in for trusted runners and warn against it on public repositories. Switching to an API key removes the cloud features.

Core features

Where it runs

The pricing doc lists only the desktop app for Free and Go, so scripted work effectively starts at Plus.

codex exec

Runs without a chat window and prints only the final message to stdout. It is read-only by default, so file-saving jobs need an explicit sandbox flag.

Image generation

Uses gpt-image-2 and accepts several reference images. All seven transparent PNG requests came back truly transparent, yet hands stayed the weak spot in our records.

Model and speed choice

Plus estimates run from 5–45 messages on GPT-6 Astra to 350–3,000 on GPT-6 Luna. Fast mode uses 2.5x included usage, Ultrafast 8x.

Code review

/review reports findings without touching your files, and @codex review works on pull requests. Local reviews use general limits. Review quality unmeasured.

AGENTS.md

Instructions load on every run, capped at 32 KiB combined by default; further files are then dropped. Ours is about 5 KB. Shorter files save usage.

Cloud and Slack

Cloud tasks run in a VM and may use more allowance than local ones. Slack setup asks for a workspace admin and a shared environment, so a solo Plus account should check first.

Built-in browser

Some computer control features are region-limited. In one project Instagram stayed blocked after a global allow, because that browser session stored a denial.

How we actually run it

  1. Read usage first

    Before a batch we read the weekly percentage from the session log Codex keeps. The figure can move between batches even when we generate nothing.

  2. Pick a working binary

    We run the CLI bundled in the ChatGPT desktop app (0.159.0 on 2026-10-13). An older npm install (0.53.0) had no executable.

  3. One command per image

    The form is codex exec -C <folder> -s workspace-write -i <reference> -- "<instruction>". With two references we pass -i twice.

  4. Fix the instruction

    Each repeats the character description word for word, then adds a scene line, ratio, background, "one image only" and a no-overwrite rule. Only the scene line changes.

  5. Run in sequence

    A script runs one image at a time. Across about 60 images each took 48–106 seconds. One failed with "Selected model is at capacity" and we reran just that image.

  6. Open every file

    We compare each image with the reference by eye. One had an extra hand, and a person decided whether to remake it. In one batch, pointing hands read as fists despite asking for a clear index finger.

  7. Log the run

    Exec runs do not appear in the app's list, so instruction, seconds and usage go in a table beside the images. The whole-percent display lets us read only roughly seven images per point.

OpenAI Codex pricing

Free

Free
  • Limited, desktop app · GPT-6 Luna · subject to rollout
  • No image generation

Go

$8/mo
  • Lightweight coding tasks · GPT-6 Luna
  • May include ads

Plus

$20/mo
  • Web, CLI, IDE, iOS and cloud integrations
  • About 15–160 local messages per five hours on GPT-6.1 Sol (official estimate)

Pro $100

$100/mo
  • No five-hour limit at present (official docs)
  • Listed in South Korea as from KRW 159,000

Pro $200

$200/mo
  • Higher limits than Pro $100
  • The plan we pay for; the KRW price is not on the first screen, so the page shows a conversion

Pro $500

$500/mo
  • Includes Ultrafast mode

Pricing checked 2026-11-06 · Official pricing page ↗

KRW prices come from the ChatGPT pricing page as served in South Korea; USD prices and the three Pro tiers come from the Codex pricing docs. The comparison uses the exchange rate noted on this page, and we could not confirm whether the KRW prices include tax.

Strengths

  • The same account continues across app, CLI, editor, iOS and cloud
  • A turn already in progress can finish past a limit, within fair use
  • Plus and Pro can buy credits past a limit instead of upgrading
  • Image generation fits in the same flow as code work
  • /status and the usage dashboard show remaining limits

What to consider

  • Image generation uses included limits 3–5 times faster on average, so a plan chosen for code work can run down quickly on images.
  • The official docs warn against using API token prices to estimate included usage. Consumption varies by model, context and tool use, and message counts are not fixed limits.
  • Signing in with an API key bills API rates separately from your subscription and removes cloud features such as code review and Slack.
  • Payments are non-refundable except where the law requires. After canceling you keep access to the end of the term, and price rises come with at least 30 days' notice. Refund eligibility depends on purchase route and location.
  • Pro's monthly API credits have eligibility conditions, expire at the end of the billing cycle if unused, and can change or end.
  • KRW prices are higher than the USD prices converted at the exchange rate. Check at checkout whether you can choose the billing currency.
  • Chats on individual plans can be used for model training by default, with an opt-out. Check the setting before pasting company code.
  • Models retire quickly: GPT-5.5 left Codex with ChatGPT sign-in on 2026-10-14. Re-measure rather than trusting cost habits from an older model.

Alternatives

  • Claude Code — Pick Claude Code if you already pay for Claude ($20, $100 or $200 a month) and want usage stated as 1x, 5x or 10x. Its free plan excludes it.
  • Cursor ↗ — Pick Cursor if you want an editor-first tool with a weekly allowance: Premium costs $30 a month, and third-party models bill separately at API list price.

FAQ

How much does Codex cost?

As of 2026-11-06: ChatGPT Free KRW 0, Go KRW 13,000 ($8) a month, Plus KRW 29,000 ($20), Pro from KRW 159,000 ($100), with $200 and $500 tiers listed in USD. Codex is included in these plans at different limits.

Can I use it for free?

Free includes limited access: per the official docs, GPT-6 Luna in the desktop app, subject to rollout, without image generation. We have not measured how far it goes.

Which should I pick: Plus or Pro?

The official docs describe Plus as a few focused sessions a week and estimate about 15–160 local messages per five hours on GPT-6.1 Sol. Pro currently has no five-hour limit. Look at Pro for long automation or heavy image work.

Who owns the code it writes?

OpenAI's terms say that, as far as the law permits, you keep your rights in input and own the output, and that other users may receive similar output. This is not legal advice.

How do cancellation and refunds work?

You can cancel at any time and keep access to the end of the term. The terms say payments are non-refundable except where the law requires; the help article says eligibility depends on purchase route and location. App Store purchases go through Apple.

What happens at a usage limit?

A turn in progress can finish within fair use. Plus and Pro can buy credits, and continuing with an API key bills API rates. You can also wait for the reset.

How does it differ from Claude Code?

The subscriptions differ: Codex comes with ChatGPT and Claude Code with Claude. Codex can also generate images in the same app. We used them on different work and never compared them on the same task, so we do not call either better.

What should I check if the CLI seems stuck?

When passing an image with -i, check that -- comes before the instruction; without it the instruction is read as an image and the command waits. In scripts, closing standard input is also safer.

Our verdict

Codex is easiest to justify when you already pay for ChatGPT, since the first test costs nothing extra. Start on Plus, run a normal week and read the dashboard before considering Pro: only credits sit between $20 and $100. In our image work, one run in about 60 failed, on capacity, and as of 2026-11-07 we still pay for Pro $200. Hands need checking by eye, and the usage display is too coarse to price a single image. We cannot tell you how its code holds up, because we used it mainly for media production. If code quality decides your purchase, test it on your own repository.

Similar tools

Logos and icons are trademarks of their companies, shown only to identify each tool. They do not imply affiliation or endorsement.

How we built this card

  • We re-read the official product and pricing pages on 2026-11-06.
  • For tools our desk uses, we record the plan, period and numbers; for tools we have not used, we say so.
  • No star ratings. This page contains no affiliate links, and every external link goes to the official site.

Sources: ChatGPT pricing · Codex pricing · Codex documentation · OpenAI Terms of Use · Canceling your ChatGPT subscription · AI 에이전트 업계 지도(ko) · AI agents market map (en) · Codex non-interactive mode · Codex image generation · Codex code review · AGENTS.md · Cloud environments · Scheduled tasks · ChatGPT in Slack · Cursor pricing