I Spent $20 On Codex And Got 24 Days Of Coding Work Done In 6 Hours - But There's A Big Catch

4 hours ago

Follow ZDNET: Add america arsenic a preferred source on Google.

ZDNET's cardinal takeaways

I spent $20 connected Codex and saw 16x productivity.
Hard usage limits trim disconnected coding successful mid-project unexpectedly.
Premium devices tin costs $800 monthly but springiness steadier results.

We've talked earlier astir OpenAI's Codex, nan AI exemplary designed specifically for programming work. While I recovered it to person immoderate capabilities, nan request to usage it only successful GitHub aliases arsenic portion of nan command-line terminal interface felt rather restrictive.

But each that changed astatine the end of August. Codex is now disposable arsenic an hold to VS Code, Cursor, and Windsurf (another AI coding IDE). It's besides disposable arsenic portion of ChatGPT Plus, OpenAI's $20/month AI plan.

Codex's integration pinch VS Code is simply a existent productivity game-changer, particularly considering nan $20/month option. I've been exploring different AI coding options, but what nan hype-machines don't show you is that it tin get very expensive, very quickly -- into nan $400-800 period range.

Also: How to usage GPT-5 successful VS Code pinch GitHub Copilot

But Codex successful nan $20/month scheme is not without its limitations. After a afloat time of use, I abruptly sewage deed pinch a 33-minute usage hold. The adjacent day, aft astir 4 hours of use, I was deed by a 90-minute hold. Then, aft different 4 hours of work, Codex unopen maine down for a week.

It feels a spot for illustration a drug. You get your first taste. You get 192 hours of activity done successful a specified 12 hours. But if you want much of that sweet, saccharine superpowered productivity, you request to salary $200 a month. It's still unclear really overmuch coding you're buying pinch nan bigger plan, but I'm honestly reasoning of going for it conscionable to proceed my streak.

The bottommost line, though, is I completed 4 projects and started a 5th successful nan clip it would person usually taken maine to possibly -- possibly -- complete one.

Now that you cognize nan halfway of my experience, fto maine return you done nan details.

AI coding devices are not cheap

In a caller discussion connected YouTube, programmers Ray Fernando and Robin Ebers shared really they each walk astatine slightest $800/month connected coding tools.

Fernando says he pays $200 a period for nan Claude Code Max 20x plan. Claude Code is nan brace programming adjunct exemplary specifically designed by Anthropic for coders, and is often preferred for its readying and reasoning capabilities.

Also: I asked AI to modify mission-critical code, and what happened adjacent haunts me

Fernando besides pays $400 per period for Cursor Ultra. Cursor is simply a fork of nan VS Code integrated improvement situation (IDE) pinch much autochthonal AI smarts. The improvement situation integrates AI capabilities passim nan interface, allowing for powerful AI-based coding assistance.

Additionally, Fernando pays $200/month for ChatGPT Pro, nan high-end type of ChatGPT pinch precocious heavy reasoning and less limits.

To Fernando and Ebers, specified expenses are worthy it for nan added productivity they're getting from nan tools. And, arsenic a matter of perspective, hiring moreover nan astir entry-level programmer would costs much than they're each spending.

Also: I retested GPT-5's coding skills utilizing OpenAI's guidance - and now I spot it moreover less

For nan moment, let's not talk whether AI is sidesplitting nan chances of entry-level programmers getting gigs. Let's conscionable attraction connected what AI tin do for master programmers who want much productivity (or much clip to sleep).

My Codex usage acquisition

Overall, my usage acquisition (when I could usage it) pinch Codex successful VS Code was rather pleasant. I worked almost wholly pinch nan chatbot interface wrong nan IDE.

My first large trial was getting Codex to rewrite my invited surface for my code. The invited surface is basal HTML and CSS, which isn't each that difficult to manus code. But I'll stock this small concealed astir me: I deeply, profoundly dislike coding successful CSS.

CSS is nan formatting "language" utilized to customize nan show of web pages. It was developed by committee, past customized and tweaked to nan constituent that nary logical programmer should beryllium capable to tolerate it. Yes, location are CSS masters. But I americium not 1 of them.

So, nan expertise to delegate that activity to nan AI was catnip to me. First, CSS is very good known and well-documented, truthful nan AI has heavy knowledge of really it works.

Also: Google's Jules AI coding instrumentality exits beta pinch superior upgrades - and much free tasks

Second, thing successful that codification is mission-critical. It tin screw up nan show each it wants and nan only harm done would beryllium to my artistic sense.

Third, each nan codification for that surface exists successful 2 files: an HTML and a CSS file. So, should nan AI render it wholly unusable, I could simply reconstruct those files from nan main branch aliases immoderate caller backup.

It did return a bunch of tries. When I first told nan AI that I wanted nan surface to look much modern and attractive, it had nary thought what to do pinch that instruction.

But erstwhile I instructed it conception by conception (i.e., make 3 achromatic boxes), it did a awesome job. So I worked pinch it, gave it 1 alteration astatine a time, waited 5-10 minutes for it to deliberation done a change, and past moved on.

Here's nan before.

Here's nan after.

The resulting type is cleaner, easier to publication and understand, and much attractive.

I besides utilized Codex to trial and debug nan mailing database signup form, which was antecedently rather buggy. Here, Codex went into nan JavaScript. Again, we iterated a bunch of times until thing came retired that worked.

Also: AI's free web scraping days whitethorn beryllium over, acknowledgment to this caller licensing protocol

It's important to support successful mind that I had to very cautiously nonstop nan AI connected what to change. I had to beryllium capable to intelligibly picture what was and wasn't working. And I had to beryllium capable to spell into my codification and nan browser console to do immoderate testing.

So, Codex is not a instrumentality for a non-programmer. A clueful quality being pinch a batch of coding and task knowledge has to beryllium astatine nan helm successful bid to make a task work.

Hard limits and reliable choices

There are, however, immoderate immense gotchas to that $20/month ChatGPT Plus scheme pinch Codex. Most coding AIs are metered based connected usage. Each clip you springiness a punctual to nan AI, immoderate number of tokens (often rather a ample number) are consumed. The higher-priced plans springiness you much tokens earlier you're trim off.

Most of nan coding devices show you nan disposable tokens you person connected your plan, and really galore you person near earlier you request to return a break. At slightest correct now, Codex successful VS Code does not.

I'd been utilizing Codex successful VS Code each time (quite literally all day). I deed 8:45 p.m. and suddenly, successful nan mediate of a reasonably analyzable and broad debugging run, I was told, "You've deed your usage limit. Upgrade to Pro aliases effort again successful 33 minutes."

Uh oh.

There was nary warning. It was virtually successful nan mediate of a coding run. So I had nary thought if it made partial changes erstwhile it stopped and indicated failure, aliases not. I didn't person entree to nan "what it changed" interface that's usually disposable aft a Codex run. My codification was successful a very indeterminate state.

This was disturbing, and it sewage worse. There is nary announcement erstwhile Codex cuts you off. As I mentioned above, it trim maine disconnected doubly more, besides without warning.

Yes, I could rotation backmost nan branch to earlier I did immoderate activity pinch Codex connected this peculiar feature, but past I'd beryllium wasting each that usage.

Speaking of usage waste, support successful mind that while programmers are paying for tokens aliases usage, nan AIs are often making boneheaded mistakes. So you're besides paying for unusable results. If my usage is immoderate indicator, astatine slightest half nan requests you make to nan coding partner AI will beryllium worthless.

Also: GPT-5 bombed my coding tests, but redeemed itself pinch codification analysis

The first clip nan prohibition hit, I didn't cognize really overmuch usage I'd get backmost erstwhile those 33 minutes are up. It turns retired nan bans are progressively longer, ever without immoderate warning, and often correct successful nan mediate of an AI coding run.

After waiting nan 33 minutes, I ran 8 much afloat coding instruction prompts, without interruption, which brought maine to nan extremity of my first time utilizing Codex this way.

The adjacent morning, aft astir 4 hours, I deed different difficult limit. This time, Codex wanted maine to hold 90 minutes. It again deed nan difficult limit successful nan mediate of 1 of its coding runs, again leaving my codification successful what seemed for illustration an chartless state.

The 90-minute activity stoppage was unsmooth because I was correct successful nan mediate of a reasonably robust travel state.

I moved onto a non-coding task for a while, until nan 90 minutes had passed.

I past had different really productive tally pinch Codex. I'd opportunity I worked for different 5 hours aliases so. But past I deed nan eventual big-time deal-killer of a ceramic wall.

Also: AI's not 'reasoning' astatine each - really this squad debunked nan manufacture hype

I'd been trim off. Codex wouldn't talk to maine for 5 days and 16 hours, unless I upgraded to nan $200/month plan.

So that's nan reliable choice. Do you limit yourself to immoderate you tin get retired of Codex connected nan $20/month Plus plan? In practice, that seems for illustration 2 days of activity -- but is that 2 days per month, aliases 2 days per week?

Because I don't person overmuch free clip to program, it mightiness beryllium workable for me. Even so, it's a buzzkill to beryllium unopen down erstwhile I still person nan clip to code. But for a master coder, that two-day on, five-day disconnected point conscionable won't work.

I've reached retired to OpenAI for immoderate specifications connected nan Plus usage limits, and really to cognize erstwhile those limits are up. Stay tuned. I'll update this erstwhile I find retired more.

And yet, what I sewage done…

Even though I was unopen retired of nan AI coding thief from Codex, I could still codification connected my own. Except, now that I've tasted what this is like, I don't cognize if I want to.

Let maine tally down what I completed successful little than 2 days of intermittent coding, successful and astir my different activities, including penning this article.

Also: I asked AI to modify mission-critical code, and what happened adjacent haunts me

It should beryllium noted that thing I did pinch nan AI was mission-critical aliases security-related code. These were each very bully capabilities to add, but thing successful these areas would origin information issues to users. That, I'm coding by hand.

1. I prettied up nan invited page

This is simply a task almost wholly HTML and CSS based, and is nan benignant of prosaic coding activity I genuinely dislike.

I'd opportunity it took maine astir 90 minutes pinch Codex. That would person been a afloat day's activity connected my own.

2. I debugged a mailing database interface

I utilized Codex to excavation done mailing database interface codification that wasn't moving properly. It would telephone retired to Mailchimp, but not succeed. I had dug into it earlier a fewer times, but ne'er pinch a reliable fix.

Using Codex, I was capable to hole it successful astir 10 minutes. That's different multi-hour savings, for a task I'd done a number of runs astatine previously.

3. I created a sub-tabs UI motor

My interface is divided into tab sections, but it's getting complex. I usage a parameterized personification interface motor to specify those tabs. Yesterday, I added a complete sub-tabs motor that creates an wholly caller personification interface component.

Not only does it supply nan caller UI elements, but it's super-modular and easy to deploy. This benignant of modular functionality add-on is rather analyzable to instrumentality and decently fastener in.

Done connected my own, it astir apt would person taken maine astatine slightest a week. It took nether 4 hours, including nan 33 infinitesimal mandatory break.

4. I added a settings backup head

I added a caller subsystem for backing up, restoring, and clearing each of nan settings successful my program. This creates an outer JSON matter record that tin beryllium easy moved, saved, and modified extracurricular of nan program's database.

Not only did I build nan codification for redeeming and restoring nan settings, but besides nan personification interface that enables tract owners to usage it.

Normally, nan UI would beryllium astir a day, and nan save/restore/reset functionality would astir apt beryllium different time aliases so. It took astir 3 hours, of which 90 minutes was a mandatory break.

5. I started a personification activity show add-on

I started activity connected thing that whitethorn go a paid add-on for my freemium product, a personification activity monitor. This is simply a very deep, feature-rich instrumentality for capturing personification and tract activity, analyzing it, and flagging suspicious activity.

I prototyped nan full personification interface, and created a first trim of nan existent logic that makes nan full point run. Unfortunately, that's arsenic acold arsenic I sewage earlier nan prohibition hammer hit. I'm trim disconnected from Codex for nan adjacent week.

Even so, getting arsenic acold arsenic I did would person taken maine astatine slightest a week for nan UI prototype, and astir apt different week for nan functional logic. I spent 4 hours, 2 of which were utilized to cautiously creation a merchandise requirements document. The different 2 hours were utilized to activity pinch nan AI to make nan code.

Bottom statement for me

I'm not exaggerating erstwhile I opportunity I sewage astir 24 days of activity done, successful 2 short coding days of astir six hours each. I completed astir 192 hours of activity successful 12 hours pinch nan thief of Codex. Talk astir a unit multiplier!

If I were coding professionally, nan costs of $200 a period for nan Pro relationship (or moreover $800 a month, arsenic Fernando and Ebers spend) would beryllium wholly worthy nan expense. But my coding is mostly to support my chops up and astir of my activity is unfastened root and free.

Also: The champion AI for coding successful 2025 (and what not to use)

I tin warrant $20 aliases moreover $40 per period for coding, but it's conscionable not thing wherever I tin warrant spending hundreds of dollars per month.

I've seen productivity benefits utilizing AI before, utilizing ChatGPT successful nan chatbot interface. But here, Codex had entree to nan full 492 record codification repo and could spot nan narration nan files had pinch each other. In different words, Codex could codification to nan wide system, not conscionable 1 module present and there.

So, while I person seen immoderate large productivity coding gains utilizing ChatGPT to thief pinch isolated routines, I've ne'er earlier seen a 16x productivity boost, which is what Codex provided maine this weekend.

But don't publication into this that AI tin make you that million-dollar app. On its own, without my changeless guidance, Codex made a ton of mistakes. It needed to beryllium instructed connected really nan various elements activity together. It needed to beryllium coached astir really to spell astir solving immoderate of nan problems. Sometimes, I had to return it wholly distant from nan attack it was utilizing and constituent it successful different directions.

In different words, it's a powerfulness tool. It's not a magical wishing well.

The powerfulness instrumentality affinity is simply a bully one. I person 4 aliases 5 different kinds of powerfulness saws successful my workshop. Each saw is bully astatine definite kinds of cutting tasks. Each saw saves tremendous clip compared to utilizing a manus saw. Each saw requires capable accomplishment to beryllium some safe and nutrient thing marginally usable.

Unless someone's an artisan (or wants to debar nan sound and sawdust produced by powerfulness saws), astir woodworkers are improbable to usage manus saws for nan bulk of their work. But conscionable having nan powerfulness saws doesn't mean nan occupation is simple. They're unit multipliers, not accomplishment replacement devices.

Also: 8 ways to constitute amended ChatGPT prompts - and get nan results you want faster

This, too, is really AI coding partners work. They are undoubtedly unit multipliers. But activity for illustration I'm trying to get done, connected a analyzable information merchandise pinch a ample installed base, still requires nan knowledge and accomplishment of a quality programmer.

It's besides present wherever nan mobility of whether AI will switch programmers becomes moreover much of a quandary. On 1 hand, moreover $1,000/month is little than nan costs of nan greenest quality coder. On nan different hand, $200-$800 per period is acold excessively rich | for astir play coders and hobbyists, arsenic good arsenic students. In different words, not only whitethorn AI brace programming trim disconnected entry-level gigs, it whitethorn besides value retired each but nan firm coders from utilizing devices for illustration Codex astatine thing resembling full-time intensity.

But damn! For nan clip I had entree to Codex this weekend, I really powered done that codification astatine warp speed. It took my activity away. I will request to deliberation overmuch connected this.

The remainder of nan story

I'm updating this 4 days aft nan items written supra and arsenic of yesterday morning, I still had 1 time and 12 hours to spell earlier I was to beryllium allowed backmost in. But I really, really wanted to spot really acold I could get connected that personification activity add-on truthful I plunked down nan $200 for a period of Pro.

I spent 8 hours yesterday and built an full tract intelligence add-on for my information software. It's wide and deep, search personification accesses, tract bots, AI bots and more. It has export capability. And it has an different activity guardian that does a immense magnitude of analytics connected tract activity to emblem imaginable problems. It includes each nan UI, fancy accordion effects, and more.

Also: Stop utilizing AI for these 9 activity tasks - here's why

That would person taken maine 8 weeks. It took maine little than 8 hours (and a batch of vexation -- nan AI tin beryllium incredibly, stubbornly stupid). It's decidedly not thing that a personification unfamiliar pinch my halfway codification could usage because nan AI kept wanting to activity successful nan incorrect place. But my gosh, nan extremity productivity is breathtaking.

So, I don't know. I want to spot if I tin build thing other into nan package successful nan adjacent 2 days. I besides want to spot if I get trim disconnected utilizing OpenAI's astir premium scheme aft a fewer days of sustained usage. Stay tuned.

Have you tried utilizing Codex aliases different AI coding assistants? Did you find nan costs worthy it compared to nan productivity boost? Would you salary $20, $200, aliases moreover $800 monthly for these tools? How do you equilibrium nan tradeoff betwixt redeeming money and keeping your coding travel uninterrupted? Let america cognize successful nan comments below.

You tin travel my day-to-day task updates connected societal media. Be judge to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.