Is Codex Free? What Actually Costs Money (2026)

TL;DR: Free Harness, Metered Brain

"Is Codex free" has two honest answers, and which one applies depends on what you mean by Codex. The command-line tool — the thing you install — is free and open source on GitHub. Nobody will ever invoice you for the CLI. But the CLI is a harness: it orchestrates requests to OpenAI's models, and those tokens are always billed to something. A subscription's credit allowance, or an API key with a card attached. There is no third option, which means a genuinely $0 Codex workflow doesn't exist.

That's not a gotcha — it's the standard shape of the whole category now, and once you see it you can price any AI coding tool in about a minute: ignore the license, ask what one real task costs in usage.

What the Free Part Actually Covers

• The CLI, fully — open source, inspectable, no feature-gated "community edition" games.
• Installation and setup — you can have it on your machine in minutes without a card.
• Nothing that touches a model — the first prompt you run needs a ChatGPT login or an API key behind it.

The practical consequence: if you already pay for ChatGPT Plus, you already have Codex. Plus at $20/month includes Codex credits, and for light-to-moderate daily use — an hour or two of real work — that allowance holds. No separate purchase, no new subscription. For the millions of people already paying for Plus, Codex arriving bundled is the closest thing to "free" this question has.

Where the Meter Starts Biting

Credits are a meter, and agentic work is the expensive unit on it. A quick "explain this function" sips; a long unsupervised run — "migrate this module, run the tests, fix what breaks" — gulps. Codex's superpower is precisely those long autonomous runs, which means the tool's best feature is also its most expensive one.

The crossover I see in practice lands around 4–5 hours of agentic work a day: that's where Plus limits start interrupting tasks mid-run, and you're choosing between watching API overage stack up in $10 increments or moving to ChatGPT Pro at $200/month. And here's the gap worth knowing before you commit: OpenAI has no middle tier. It's $20, then $200. Claude Code offers a $100 Max step in between, which is why a 3–4 hour/day user is arguably better priced on Claude — I ran that comparison properly in the full Codex pricing breakdown.

Update (September 2026): that gap has closed. OpenAI's Codex pricing page, read on September 8, 2026, lists Pro from $100/month (5x the Plus limits) alongside the $200 tier (20x), a Go plan at $8, and credits you can buy on Plus when you hit the wall instead of upgrading. What the wall actually is, a rolling five-hour allowance per model with a weekly cap that may apply, is in Codex usage limits explained.

The Pay-As-You-Go Route

Skip the subscription entirely and Codex bills against an OpenAI API key. From the tier math in the pricing guide: roughly $5–$15/month for about an hour a day, $30–$60 for 3–4 hours, $80–$250 for full-time agentic work. The API route makes sense at the edges — very light dabbling where $20/month would be wasted, or team setups where you want one billing surface. For everyone in the middle, the bundled subscription credits are simpler and usually cheaper.

How to Decide in One Afternoon

Don't price this from a blog post — including mine. Install the CLI (free), point it at your actual codebase, and give it one task you were dreading. Watch two numbers: how much of your Plus allowance the task consumed (that's your tier forecast), and how much of the diff you'd actually ship. The first number tells you whether you're a $20 user or a $200 one. The second tells you something more important.

Both numbers depend heavily on how you work it — scope, approval mode, and whether you delegate or supervise. If you want the allowance to stretch, start with how to use Codex.

The Part That Isn't on Any Pricing Page

Because here's what I owe you as the person who fixes these codebases for a living: the subscription is the cheapest line item in AI-assisted development. Codex's long unsupervised runs are genuinely impressive — and unreviewed output from them is exactly how projects end up in my rescue queue: apps that demo beautifully and are missing the boring 20% — authorization checks, input validation, webhook verification. Fixing that after launch runs $3,500–$15,000 on my real invoices. Against that, the difference between free, $20, and $200 is noise. The compounding version of this problem is what I call Vibe Debt, and "free" tools are where it usually starts, because free quietly encourages accepting more and reviewing less.

My Honest Recommendation

If you're Plus-subscribed already: use Codex today, pay nothing extra, and see whether your usage ever threatens the limits. If you're starting from zero: $20 Plus is the right first dollar, and don't pre-buy the $200 tier until the meter tells you you're that user. If you're comparing tools before committing, the three-way against Cursor and Claude Code is the comparison that actually settles it — and my six-month Codex review covers what the meter doesn't.

Whatever tier you land on: before real users or real money touch what Codex built, run it against my free 20-point vibe coding security checklist — the same audit I run on paid rescue work. Already live and getting a bad feeling? Book a strategy call and I'll give you a straight answer.

Update: the $200 Pro tier is open again

Update (September 20, 2026): OpenAI paused new sign-ups and upgrades to Pro 20x ($200/month) on September 10, 2026, after demand for GPT-6 Astra (launched September 3) outran capacity. September 29, 2026: the pause is over. New Pro 200 subscriptions are open again, OpenAI canceled the allowance cut it had planned on October 29, and Pro 500 ($500/month) now sits on top. Moving up from a free or Plus setup, I'd still try Pro at $100 before either.