OpenAI Codex Pricing (2026): What It Actually Costs, Tier by Tier

TL;DR: The CLI Is Free. The Question Is Which Meter You Run.

OpenAI's pricing page tells you what the tiers cost. It doesn't tell you which one your workload lands in. I've run Codex daily for six months alongside Claude Code — same repos, same kinds of tasks — so the numbers below are what the meters actually read, not what the marketing implies.

There are exactly two ways to pay for Codex, and most developers should start with the one they may already have:

• Bundled with ChatGPT: Plus ($20/month) includes Codex credits — enough for light-to-moderate daily use. Pro ($100/month for 5x the Plus limits, $200 for 20x) effectively removes the meter for one person; the $200 tier is open to new sign-ups again as of September 29, 2026, and a $500 tier sits above it.
• API pay-as-you-go: usage billed to your OpenAI key. Fine for automation and CI. Expensive as a daily-driver habit.

Codex Pricing by Workload (August 2026)

Those API bands mirror what I published in my six-month Codex review — they've held steady through 2026 even as the model lineup underneath shifted.

The Middle Tier: $20, $100, Then $200

Here's the structural thing about Codex pricing that surprises people coming from the Claude side. Anthropic sells a ladder: Pro $20, Max $100, Max $200. Until April 2026 OpenAI sold a cliff: Plus $20, then Pro $200, nothing between.

Update (September 2026): the cliff now has a step in it. OpenAI's own Codex pricing page, read on September 8, 2026, lists Pro from $100/month in two tiers, 5x ($100) and 20x ($200) the Plus rate limits, plus a Go tier at $8/month for lightweight coding and a Business plan at $20 per user per month billed annually ($25 monthly, two-user minimum). The ladder below still describes the tiers as they were when this post was written; the limits themselves, the five-hour window and the weekly cap, are in Codex usage limits explained.

If you're a 3–4 hour/day agentic user, you're the person the $100 step was built for. Plus will mostly cover you, but on heavy days you'll hit limits mid-task. Your options at that point: eat the occasional limit, run API overage on top of Plus (fine until it quietly becomes $60/month), or move to Pro 5x at $100 and stop thinking about it. The $200 tier is for the 5-plus-hour-a-day person, and it's open to new sign-ups again, at the higher allowance now that OpenAI canceled its planned cut on October 29, 2026. There's no wrong answer, but decide it once instead of drifting into the expensive version $10 at a time — the same drift I documented in my Claude Code pricing breakdown, where API auto-reload was the trap.

Codex vs Claude Code vs Cursor on Price

Entry tier: identical — $20/month gets you real daily use of either terminal agent or Cursor Pro. Top tier: also basically identical at $200, and both are open to new buyers as of September 29, 2026, though OpenAI now puts a $500 tier above its $200 one. The middle band is where the shopping happens: ChatGPT Pro 5x and Claude Max 5x are both $100, and Cursor's $60 Pro+ undercuts both for the heavy-but-not-full-time user. If that's you and you're price-sensitive, that's a legitimate reason to pick a different primary tool — the capability gaps are smaller than the pricing gaps.

Update, October 16, 2026: Cursor's entry price moved. Since October 12, cursor.com/pricing sells The X Subscription to new buyers instead of the $20 Pro plan. Plus at $8 runs Grok models only, and Premium at $30 is the first tier that runs Claude or GPT models in Cursor, with that usage billed on top at API list prices. So for a new buyer who wants OpenAI models in an agent, Codex inside ChatGPT Plus at $20 is now the cheaper way in. The top tiers still meet at $200 (Cursor Ultra). Existing Cursor Pro subscribers keep their plan. Every new tier: Cursor pricing.

One more Codex-specific advantage worth naming: if your company already pays for ChatGPT anyway, Codex is the only one of the three that's free at the margin. That single fact decides the tooling question for a lot of teams before any benchmark does. If the free question is the whole reason you're here, I answered it properly — what free covers and where the meter starts — in Is Codex free?

What I'd Pick, By Situation

• Already on ChatGPT Plus: just use Codex. You're paying for it either way. Upgrade only when limits actually interrupt you.
• Choosing a first AI coding tool at $20: read the head-to-head — pick on workflow fit, not price; they cost the same here.
• 4–5+ hours/day of agentic work: ChatGPT Pro, $100 for 5x the Plus limits, with $200 and $500 tiers above it (the $200 tier reopened to new sign-ups by September 29, 2026). The bundled credits make the meter irrelevant, which changes how ambitiously you use the tool. Whether that tier is actually worth it for you: is ChatGPT Pro worth it.
• Automation and CI: API key, scoped tasks, budget alerts on. Same rules as any metered compute.

Whichever meter you land on, the variable that actually moves your bill is how you work — scope, approval mode, and how much you delegate rather than supervise. That's covered separately in how to use Codex.

Codex Pro, and Plus vs Pro for Codex

There is no separate Codex Pro product. "Codex Pro" means running Codex on a ChatGPT Pro plan, and as of September 29, 2026 OpenAI's Codex pricing page lists Pro at $100, $200 or $500 a month. OpenAI paused new $200 sign-ups on September 10; its help center, read September 29, said Pro 200 was open to new subscriptions again. On October 29 OpenAI canceled the lower allowance it had planned, so every Pro 200 subscriber gets the higher one. For a new buyer the first choice is still Plus at $20 or Pro at $100.

What Pro buys in Codex now is the five-hour window going away. OpenAI's page says Pro plans currently have no five-hour limit, while Plus gets roughly 15 to 150 GPT-6 Sol local messages per window; weekly limits can still apply to both. Codex is included on every plan, Free and Go too; the tiers change how much you can run, not what you can run.

Also new on DevDay (September 29, 2026): the $500 plan is the only self-serve plan with Astra Ultrafast, a faster GPT-6 Astra mode that draws included usage at 8x the Standard rate. On Plus, the new GPT-6.1 Sol is estimated at 15 to 160 local messages per five hours. GPT-5.5 leaves Codex on every plan on October 14, 2026. For almost everyone reading a pricing page, the $500 tier changes nothing: the decision is still Plus at $20 or Pro at $100.

Who should pay for Pro: someone whose /status shows Plus running dry most working days on Sol. If it happens once a week, Plus lets you buy credits without changing plan, which costs less than $80 more a month. A Business Premium seat is the other $100 route: it gets 5x a Standard seat's usage with no 5-hour limit, and adds admin controls and no training on your data by default.

The Bill Nobody Prices: What the Agent Ships

The subscription is the small number. The expensive part of any AI coding agent is unreviewed output making it to production — I've rebuilt enough AI-generated codebases to say the cleanup quote is routinely 100x the subscription that produced the mess. If Codex (or any agent) is shipping features customers touch, run my free 20-point vibe coding security checklist against the repo before launch. And if you're past that point and things are wobbling, book a strategy call — first one's free.