Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- I spent $20 on Codex and saw 16x productivity.
- Hard usage limits cut off coding in mid-project unexpectedly.
- Premium tools can cost $800 monthly but give steadier results.
We’ve talked before about OpenAI’s Codex, the AI model designed specifically for programming work. While I found it to have some capabilities, the need to use it only in GitHub or as part of the command-line terminal interface felt quite restrictive.
But all that changed at the end of August. Codex is now available as an extension to VS Code, Cursor, and Windsurf (another AI coding IDE). It’s also available as part of ChatGPT Plus, OpenAI’s $20/month AI plan.
Codex’s integration with VS Code is a true productivity game-changer, especially considering the $20/month option. I’ve been exploring other AI coding options, but what the hype-machines don’t tell you is that it can get very expensive, very quickly — into the $400-800 month range.
Also: How to use GPT-5 in VS Code with GitHub Copilot
But Codex in the $20/month plan is not without its limitations. After a full day of use, I suddenly got hit with a 33-minute usage hold. The next day, after about four hours of use, I was hit by a 90-minute hold. Then, after another four hours of work, Codex shut me down for a week.
It feels a bit like a drug. You get your first taste. You get 192 hours of work done in a mere 12 hours. But if you want more of that sweet, sweet superpowered productivity, you need to pay $200 a month. It’s still unclear how much coding you’re buying with the bigger plan, but I’m honestly thinking of going for it just to continue my streak.
The bottom line, though, is I completed four projects and started a fifth in the time it would have normally taken me to maybe — maybe — complete one.
Now that you know the core of my experience, let me take you through the details.
AI coding tools are not cheap
In a recent discussion on YouTube, programmers Ray Fernando and Robin Ebers shared how they each spend at least $800/month on coding tools.
Fernando says he pays $200 a month for the Claude Code Max 20x plan. Claude Code is the pair programming assistant model specifically designed by Anthropic for coders, and is often preferred for its planning and reasoning capabilities.
Also: I asked AI to modify mission-critical code, and what happened next haunts me
Fernando also pays $400 per month for Cursor Ultra. Cursor is a fork of the VS Code integrated development environment (IDE) with more native AI smarts. The development environment integrates AI capabilities throughout the interface, allowing for powerful AI-based coding assistance.
Additionally, Fernando pays $200/month for ChatGPT Pro, the high-end version of ChatGPT with advanced deep reasoning and fewer limits.
To Fernando and Ebers, such expenses are worth it for the added productivity they’re getting from the tools. And, as a matter of perspective, hiring even the most entry-level programmer would cost more than they’re each spending.
Also: I retested GPT-5’s coding skills using OpenAI’s guidance – and now I trust it even less
For the moment, let’s not discuss whether AI is killing the chances of entry-level programmers getting gigs. Let’s just focus on what AI can do for professional programmers who want more productivity (or more time to sleep).
My Codex usage experience
Overall, my usage experience (when I could use it) with Codex in VS Code was quite pleasant. I worked almost entirely with the chatbot interface inside the IDE.
My first big test was getting Codex to rewrite my welcome screen for my code. The welcome screen is basic HTML and CSS, which isn’t all that hard to hand code. But I’ll share this little secret about me: I deeply, deeply dislike coding in CSS.
CSS is the formatting “language” used to customize the display of web pages. It was developed by committee, then customized and tweaked to the point that no rational programmer should be able to tolerate it. Yes, there are CSS masters. But I am not one of them.
So, the ability to delegate that work to the AI was catnip to me. First, CSS is very well known and well-documented, so the AI has deep knowledge of how it works.
Also: Google’s Jules AI coding tool exits beta with serious upgrades – and more free tasks
Second, nothing in that code is mission-critical. It can screw up the display all it wants and the only harm done would be to my aesthetic sense.
Third, all the code for that screen exists in two files: an HTML and a CSS file. So, should the AI render it completely unusable, I could simply restore those files from the main branch or any recent backup.
It did take a bunch of tries. When I first told the AI that I wanted the screen to look more modern and attractive, it had no idea what to do with that instruction.
But when I instructed it section by section (i.e., make three white boxes), it did a great job. So I worked with it, gave it one change at a time, waited 5-10 minutes for it to think through a change, and then moved on.
Here’s the before.
Here’s the after.
The resulting version is cleaner, easier to read and understand, and more attractive.
I also used Codex to test and debug the mailing list signup form, which was previously quite buggy. Here, Codex went into the JavaScript. Again, we iterated a bunch of times until something came out that worked.
Also: AI’s free web scraping days may be over, thanks to this new licensing protocol
It’s important to keep in mind that I had to very carefully direct the AI on what to change. I had to be able to clearly describe what was and wasn’t working. And I had to be able to go into my code and the browser console to do some testing.
So, Codex is not a tool for a non-programmer. A clueful human being with a lot of coding and project knowledge has to be at the helm in order to make a project work.
Hard limits and tough choices
There are, however, some huge gotchas to that $20/month ChatGPT Plus plan with Codex. Most coding AIs are metered based on usage. Each time you give a prompt to the AI, some number of tokens (often quite a large number) are consumed. The higher-priced plans give you more tokens before you’re cut off.
Most of the coding tools show you the available tokens you have on your plan, and how many you have left before you need to take a break. At least right now, Codex in VS Code does not.
I’d been using Codex in VS Code all day (quite literally all day). I hit 8:45 p.m. and suddenly, in the middle of a fairly complex and comprehensive debugging run, I was told, “You’ve hit your usage limit. Upgrade to Pro or try again in 33 minutes.”
Uh oh.
There was no warning. It was literally in the middle of a coding run. So I had no idea if it made partial changes when it stopped and indicated failure, or not. I didn’t have access to the “what it changed” interface that’s normally available after a Codex run. My code was in a very indeterminate state.
This was disturbing, and it got worse. There is no notice when Codex cuts you off. As I mentioned above, it cut me off twice more, also without warning.
Yes, I could roll back the branch to before I did any work with Codex on this particular feature, but then I’d be wasting all that usage.
Speaking of usage waste, keep in mind that while programmers are paying for tokens or usage, the AIs are often making boneheaded mistakes. So you’re also paying for unusable results. If my usage is any indicator, at least half the requests you make to the coding partner AI will be worthless.
Also: GPT-5 bombed my coding tests, but redeemed itself with code analysis
The first time the ban hit, I didn’t know how much usage I’d get back when those 33 minutes are up. It turns out the bans are progressively longer, always without any warning, and often right in the middle of an AI coding run.
After waiting the 33 minutes, I ran eight more full coding instruction prompts, without interruption, which brought me to the end of my first day using Codex this way.
The next morning, after about four hours, I hit another hard limit. This time, Codex wanted me to wait 90 minutes. It again hit the hard limit in the middle of one of its coding runs, again leaving my code in what seemed like an unknown state.
The 90-minute work stoppage was rough because I was right in the middle of a fairly robust flow state.
I moved onto a non-coding project for a while, until the 90 minutes had passed.
I then had another really productive run with Codex. I’d say I worked for another five hours or so. But then I hit the ultimate big-time deal-killer of a brick wall.
Also: AI’s not ‘reasoning’ at all – how this team debunked the industry hype
I’d been cut off. Codex wouldn’t talk to me for 5 days and 16 hours, unless I upgraded to the $200/month plan.
So that’s the tough choice. Do you limit yourself to whatever you can get out of Codex on the $20/month Plus plan? In practice, that seems like two days of work — but is that two days per month, or two days per week?
Because I don’t have much free time to program, it might be workable for me. Even so, it’s a buzzkill to be shut down when I still have the time to code. But for a professional coder, that two-day on, five-day off thing just won’t work.
I’ve reached out to OpenAI for some details on the Plus usage limits, and how to know when those limits are up. Stay tuned. I’ll update this when I find out more.
And yet, what I got done…
Even though I was shut out of the AI coding help from Codex, I could still code on my own. Except, now that I’ve tasted what this is like, I don’t know if I want to.
Let me run down what I completed in less than two days of intermittent coding, in and around my other activities, including writing this article.
Also: I asked AI to modify mission-critical code, and what happened next haunts me
It should be noted that nothing I did with the AI was mission-critical or security-related code. These were all very nice capabilities to add, but nothing in these areas would cause security issues to users. That, I’m coding by hand.
1. I prettied up the welcome page
This is a task almost entirely HTML and CSS based, and is the kind of prosaic coding activity I truly dislike.
I’d say it took me about 90 minutes with Codex. That would have been a full day’s work on my own.
2. I debugged a mailing list interface
I used Codex to dig through mailing list interface code that wasn’t working properly. It would call out to Mailchimp, but not succeed. I had dug into it before a few times, but never with a reliable fix.
Using Codex, I was able to fix it in about 10 minutes. That’s another multi-hour savings, for a project I’d done a number of runs at previously.
3. I created a sub-tabs UI engine
My interface is divided into tab sections, but it’s getting complex. I use a parameterized user interface engine to specify those tabs. Yesterday, I added a complete sub-tabs engine that creates an entirely new user interface component.
Not only does it provide the new UI elements, but it’s super-modular and easy to deploy. This kind of modular functionality add-on is quite complex to implement and properly lock in.
Done on my own, it probably would have taken me at least a week. It took under four hours, including the 33 minute mandatory break.
4. I added a settings backup manager
I added a new subsystem for backing up, restoring, and clearing all of the settings in my program. This creates an external JSON text file that can be easily moved, saved, and modified outside of the program’s database.
Not only did I build the code for saving and restoring the settings, but also the user interface that enables site owners to use it.
Normally, the UI would be about a day, and the save/restore/reset functionality would probably be another day or so. It took about three hours, of which 90 minutes was a mandatory break.
5. I started a user activity monitor add-on
I started work on something that may become a paid add-on for my freemium product, a user activity monitor. This is a very deep, feature-rich tool for capturing user and site activity, analyzing it, and flagging suspicious activity.
I prototyped the entire user interface, and created a first cut of the actual logic that makes the whole thing run. Unfortunately, that’s as far as I got before the ban hammer hit. I’m cut off from Codex for the next week.
Even so, getting as far as I did would have taken me at least a week for the UI prototype, and probably another week for the functional logic. I spent four hours, two of which were used to carefully design a product requirements document. The other two hours were used to work with the AI to generate the code.
Bottom line for me
I’m not exaggerating when I say I got about 24 days of work done, in two short coding days of about six hours each. I completed roughly 192 hours of work in 12 hours with the help of Codex. Talk about a force multiplier!
If I were coding professionally, the cost of $200 a month for the Pro account (or even $800 a month, as Fernando and Ebers spend) would be totally worth the expense. But my coding is mostly to keep my chops up and most of my work is open source and free.
Also: The best AI for coding in 2025 (and what not to use)
I can justify $20 or even $40 per month for coding, but it’s just not something where I can justify spending hundreds of dollars per month.
I’ve seen productivity benefits using AI before, using ChatGPT in the chatbot interface. But here, Codex had access to the entire 492 file code repo and could see the relationship the files had with each other. In other words, Codex could code to the overall system, not just one module here and there.
So, while I have seen some big productivity coding gains using ChatGPT to help with isolated routines, I’ve never before seen a 16x productivity boost, which is what Codex provided me this weekend.
But don’t read into this that AI can make you that million-dollar app. On its own, without my constant guidance, Codex made a ton of mistakes. It needed to be instructed on how the various elements work together. It needed to be coached about how to go about solving some of the problems. Sometimes, I had to take it completely away from the approach it was using and point it in other directions.
In other words, it’s a power tool. It’s not a magical wishing well.
The power tool analogy is a good one. I have four or five different kinds of power saws in my workshop. Each saw is good at certain kinds of cutting tasks. Each saw saves tremendous time compared to using a hand saw. Each saw requires enough skill to be both safe and produce something marginally usable.
Unless someone’s an artisan (or wants to avoid the noise and sawdust produced by power saws), most woodworkers are unlikely to use hand saws for the bulk of their work. But just having the power saws doesn’t mean the job is simple. They’re force multipliers, not skill replacement devices.
Also: 8 ways to write better ChatGPT prompts – and get the results you want faster
This, too, is how AI coding partners work. They are undoubtedly force multipliers. But work like I’m trying to get done, on a complex security product with a large installed base, still requires the knowledge and skill of a human programmer.
It’s also here where the question of whether AI will replace programmers becomes even more of a quandary. On one hand, even $1,000/month is less than the cost of the greenest human coder. On the other hand, $200-$800 per month is far too rich for most weekend coders and hobbyists, as well as students. In other words, not only may AI pair programming cut off entry-level gigs, it may also price out all but the corporate coders from using tools like Codex at anything resembling full-time intensity.
But damn! For the time I had access to Codex this weekend, I really powered through that code at warp speed. It took my breath away. I will need to think much on this.
The rest of the story
I’m updating this four days after the items written above and as of yesterday morning, I still had one day and 12 hours to go before I was to be allowed back in. But I really, really wanted to see how far I could get on that user activity add-on so I plunked down the $200 for a month of Pro.
I spent eight hours yesterday and built an entire site intelligence add-on for my security software. It’s wide and deep, tracking user accesses, site bots, AI bots and more. It has export capability. And it has an unusual activity guardian that does a huge amount of analytics on site activity to flag potential problems. It includes all the UI, fancy accordion effects, and more.
Also: Stop using AI for these 9 work tasks – here’s why
That would have taken me eight weeks. It took me less than eight hours (and a lot of frustration — the AI can be incredibly, stubbornly stupid). It’s definitely not something that a person unfamiliar with my core code could use because the AI kept wanting to work in the wrong place. But my gosh, the end productivity is breathtaking.
So, I don’t know. I want to see if I can build something else into the software in the next two days. I also want to see if I get cut off using OpenAI’s most premium plan after a few days of sustained usage. Stay tuned.
Have you tried using Codex or other AI coding assistants? Did you find the cost worth it compared to the productivity boost? Would you pay $20, $200, or even $800 monthly for these tools? How do you balance the tradeoff between saving money and keeping your coding flow uninterrupted? Let us know in the comments below.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
Leave a Reply