If you are buying for a company, the comparison is already over: Codex is purchasable and Jules is not. Google states that paid Jules plans run through a Google AI subscription that is only available on individual Google Accounts ending in @gmail.com, so a Workspace account on your own domain cannot upgrade at all. If you are buying for yourself, Jules is the better deal by a wide margin, because its free tier gives you 15 tasks per rolling 24 hours at no cost and about $19.99 raises that to 100, while Codex asks $20 before you get past a rate-limited free tier.
That is the short version. The longer version is more interesting, because these two products meter completely different things and most comparisons quietly pretend they do not. Google counts whole tasks in a rolling 24 hour window. OpenAI counts messages in a rolling five hour window. A "task" and a "message" are not the same unit and no exchange rate exists, so anyone who tells you Jules is cheaper per unit of work is guessing. What you can compare honestly is the shape of each ceiling and what happens when you hit it.
Every figure below was read from Google's and OpenAI's own documentation on September 4, 2026.
Jules vs Codex, side by side
| Google Jules | OpenAI Codex | |
|---|---|---|
| Free tier | 15 tasks per rolling 24h, 3 concurrent | Included with ChatGPT Free, rate limited |
| Entry paid tier | About $19.99 (Google AI Pro): 100 tasks, 15 concurrent | $20 (ChatGPT Plus): 10 to 100 Sol messages per 5h |
| Top tier | $100 or $200 (Google AI Ultra): 300 tasks, 60 concurrent | $200 (Pro 20x): 200 to 2,000 Sol messages per 5h |
| What the meter counts | Whole tasks | Messages, and cloud chats share the same allowance |
| Window | Rolling 24 hours | Rolling 5 hours, with weekly limits also possible |
| Can a company buy the paid tier? | No. Personal @gmail.com accounts only | Yes. Business from $20 per user billed annually |
| Model on the free tier | Gemini 2.5 Pro, a generation behind | Same model family, smaller allowance |
| Repository hosts | GitHub only | Local repositories, plus GitHub for cloud tasks |
| Where the work runs | Google cloud VM, always | Your terminal, or Codex cloud |
| Product status | Public beta | Generally available |
The difference that decides it: who is allowed to buy
Google's limits page contains one sentence that outranks every number in the table above. Paid Jules plans are accessed through a Google AI Plans subscription, "currently available only for individual Google Accounts (ending in @gmail.com)". Business power users who want more access are pointed at an interest form. There is no enterprise SKU, no seat count, no admin console, and no way to put a Jules licence on a company account.
Play that out. Your engineers sign in to everything with an address on your domain, managed in Workspace. They can use Jules on those accounts, but they cannot upgrade. So the free tier is not the first rung of a ladder, it is the roof: fifteen tasks a rolling day, three at a time, on Gemini 2.5 Pro rather than the current model. The only workaround is asking developers to run company code through personal gmail accounts, which is a conversation that ends the moment anyone from security joins it. A licence that cannot be issued to the company never reaches a purchase order in the first place, which is where most of these evaluations actually die.
OpenAI has the opposite posture. ChatGPT Business starts at $20 per user billed annually, or $25 monthly, with a minimum of two users, and it comes with admin controls, SSO and a stated default of no training on your business data. It is bought the way software is normally bought. We break down what those seats do and do not include on our Codex CLI pricing and usage limits page.
How much does Google Jules cost?
Nothing, for most people. Jules has no price of its own, so what you pay is a Google AI subscription and Jules limits rise with it: free at 15 daily tasks, about $19.99 for Pro at 100 daily tasks and 15 concurrent, and $100 or $200 for Ultra at 300 daily tasks and 60 concurrent. At four tasks a working day, which is a normal week for one developer, the free tier already covers you.
There is one arithmetic trap in that ladder and it costs $100 a month. Google has sold Ultra at two prices since May 2026, and the $200 option advertises four times the usage limits of the $100 one. Jules publishes a single "Jules in Ultra" row: 300 tasks, 60 concurrent, and that row is the same on both. The multiplier applies to Gemini usage, not to Jules' documented task limits. If Jules is your reason for looking at Ultra, the $200 tier doubles your cost per task and buys you nothing. The full ladder, including cost per task at realistic usage, is on our Google Jules pricing breakdown.
How the two ceilings feel when you hit them
This is where the units stop being an abstraction. A Jules ceiling is calm: the new task button goes grey, a tooltip explains why, and your existing tasks, diffs and history stay readable. Capacity comes back gradually through the following day because the window rolls rather than resetting at midnight. You lose throughput, not work in progress.
A Codex ceiling is sharper, because the five hour window is short enough to interrupt a single afternoon and because local terminal messages and cloud chats draw on the same allowance. A batch of cloud tasks you dispatched at ten can be the reason your terminal stops answering at three, and cloud chats run on Sol, the most expensive model in the family. The upside is that it resets four or five times a day rather than once, so you are never blocked for long.
Both are ceilings you cannot buy past in the moment. Neither vendor sells overage on the consumer tiers, so when you run out you wait. That is worth knowing before you plan a release week around either one.
The model split nobody mentions
Google lists Gemini 2.5 Pro for the free Jules tier and higher or priority access to the latest model, starting with Gemini 3 Pro, for Pro and Ultra. Combine that with the gmail-only rule and the result is genuinely awkward: the account type every US business is stuck on is also the one held a model generation back. On a small bug fix you will not notice. On a multi-file refactor with a test suite to satisfy, you will.
Codex splits differently. Every plan reaches the same model family and what changes is how many messages you get, which is why switching model matters more than switching plan there. On a $20 Plus seat, Sol allows 10 to 100 local messages per five hour window and Luna allows 250 to 2,000. Moving routine work to the cheaper model multiplies your allowance by more than upgrading two tiers would, for nothing.
Which one should you actually buy?
A solo developer with a personal gmail and a GitHub repo: Jules, on the free tier, and do not pay anyone until you are genuinely hitting fifteen tasks a day. This is the clearest recommendation in the whole comparison and it costs you nothing.
A heavy individual user: Google AI Pro at about $19.99 is 6.7 times the tasks and five times the concurrency for the price of one editor seat, and it moves you onto the current model. Skip Ultra unless you are dispatching more than a hundred tasks a day, and skip the $200 Ultra entirely if Jules is the reason.
A developer who lives in the terminal: Codex, because Jules has no local mode at all. Every Jules task runs in a Google VM, which is the point of the product and also its limitation. If you want an agent editing the files in front of you, this is not a close call.
Anyone on GitLab: Neither, for now. Google states that Jules can only access repositories you explicitly allow through GitHub, with other version control systems promised for the future.
A team buying for a company: Codex, or something else, but not Jules, until Google ships a business upgrade path. This is not a judgment about quality. It is that one of them has a checkout for organizations and the other has a form.
Where a third shape fits
Both of these ask you to adopt a vendor's whole account model to get an agent: a Google consumer subscription attached to a personal address, or a ChatGPT workspace. That is fine when you already live in one of those ecosystems and awkward when you do not.
Agentcode connects to a GitHub or GitLab repository, takes a described task, plans it, edits the code, runs your test suite and opens a pull request for a human to read. It never merges on its own, so your branch protection rules and reviewers stay exactly where they are, and the licence belongs to the company rather than to somebody's personal email. It costs $29 a month, flat, billed yearly, with no daily task ceiling and no rolling window to wait out. That is a narrower product than either of these and we would rather say so. If you are still shortlisting, every AI coding tool priced side by side and Codex against GitHub Copilot cover the field honestly, including the rows where we lose.