Argon Nailed the Pelican Test. Then It Burned My Weekly Quota.

One questionable subscription, a surprisingly talented bird, and a weekly AI quota that didn't survive the experiment.

There are certain things you should never do within five minutes of waking up.

Making financial decisions is pretty high on that list.

A few days ago, I opened my eyes, grabbed my phone, and saw the news: Google's Gemini 4 Argon was finally available.

I'd been wanting to try Argon for a while. The benchmarks looked impressive, the early reactions were promising, and I was curious to see what it could actually do.

Then I spotted two dangerous words on Google's subscription page:

Special Discount.

Something happens to the human brain when it sees a discount. Five seconds earlier, I wasn't sure I needed the product. Suddenly, not buying it felt financially irresponsible.

So, with approximately three functioning brain cells, I committed to a full year of Gemini Pro.

Payment successful.

Back to sleep.

When I woke up properly, I updated Antigravity, ready to enjoy my newly purchased productivity.

Opened the model selector.

Hmm.

Where's Argon?

A little digging revealed the catch: Argon was available in Antigravity, all right. Just for Ultra subscribers.

I had Pro.

To be fair, Google hadn't lied. Argon was available. Just not available to me, at least not where I wanted to use it.

It felt like booking a hotel for its gorgeous rooftop infinity pool, changing into your swimsuit, riding the elevator all the way up, and discovering the pool was reserved for penthouse guests.

The room was nice, though. Lovely bathrobe.

Well, at least I wouldn't have to debate whether to subscribe to Gemini Pro for the rest of the year.

That decision had been taken care of.

Enter the Pelican

Fortunately, Google moved faster than my buyer's remorse.

A couple of days later, Antigravity popped up a notification. Argon was now available to Pro subscribers.

I opened the app, checked the model list, and there it was.

All was forgiven.

It's embarrassing how little it takes to make me happy sometimes.

The timing was perfect, too. I'd already burned through my Codex weekly usage and was waiting for the reset. My usual workhorse, GPT-6.1 Sol XH, was temporarily out of commission.

Time to put the new hire to work.

I started with the pelican test.

And honestly? Not bad at all.

There was a cheerful pelican riding a green bicycle, wearing a blue hat and a red scarf, with a couple of fish sitting in the front basket.

The composition worked. The details were surprisingly coherent. No mysterious extra limbs, no bicycle frame that defied the laws of physics.

The bird looked like it had just finished a delivery and was heading to the beach for its next order.

I stared at it for a moment.

Maybe Google really had something here.

The pelican looked exceptionally pleased with itself.

Here was a bird successfully riding a bicycle, and I was still sitting at my desk.

Clearly, I needed to work harder.

So naturally, I decided to stop working.

I had Argon now.

One Goal. Zero Quota.

My next idea seemed perfectly reasonable at the time.

If Argon was this capable, why not give it something substantial? A proper development task, with existing context to understand, decisions to make, and actual work to deliver.

I spent some time writing the instructions, selected Argon High, and launched the whole thing with /goal.

Then I let it run.

And run.

And run some more.

To its credit, it certainly looked busy.

Logs were flying past. Files were being examined. Changes were happening. The whole thing had that reassuring air of serious engineering work.

I started feeling optimistic.

Maybe this was it. Maybe I'd finally found another agent I could trust with the heavy lifting.

Eventually, it finished.

I opened the result.

Well.

You know that polite email you send when a contractor delivers something you don't want to approve, but you're trying very hard to remain a civilized human being?

"Thanks for all the work. Let's revisit the direction."

That one.

Except there was no contractor to email. Just me, staring at a result I'd probably have to redo myself.

Still, these things happen. Models make mistakes. No big deal.

Then I checked the usage meter.

Weekly quota exhausted.

Oh.

Oh, that's lovely.

Codex was waiting for its weekly reset. Now Argon was waiting for its weekly reset.

I'd somehow managed to turn two cutting-edge AI coding tools into two very expensive waiting rooms.

All that money spent on productivity, and I'd successfully arranged paid time off.

For the AIs.

Not for me.

Later, I looked into how Antigravity's /goal mode actually works.

Turns out, there's no fixed iteration cap. The agent is supposed to keep going, checking and correcting its own work, until it considers the objective complete or you tell it to stop.

Which is genuinely useful when the task has clear, verifiable requirements.

My task, unfortunately, involved quite a bit of judgment. Getting something to work and getting something right aren't always the same thing.

In hindsight, I'd handed an open-ended assignment to an extremely enthusiastic employee who had no concept of billable hours.

And then I was surprised by the bill.

That's partly on me.

Next time I use /goal, I'm checking the remaining quota first.

Otherwise, I can easily imagine asking it to fix a minor issue, coming back three hours later, and finding an entirely restructured codebase.

The original bug would still be there, naturally.

So, Is Argon Any Good?

I'll keep using it.

Obviously.

I've already paid for the year. Quitting now would mean admitting that my half-asleep financial judgment might not have been flawless.

But I'm not handing it another large, loosely defined task just yet.

For now, GPT-6.1 Sol XH remains my preferred choice for serious development work. It handles complex context and larger engineering tasks more reliably in my experience.

It still makes mistakes, of course. But when it does, I can usually follow its reasoning well enough to understand what went wrong.

That's worth more than it sounds.

Opus 5.5 High is excellent, too. Some of its work is genuinely impressive.

The problem is how quickly it burns through usage.

Using Opus feels like taking a taxi in Zurich on a Nairobi paycheck.

The driver is excellent. The ride is smooth. But every time the meter ticks up, I start reconsidering whether I really needed to go anywhere in the first place.

Grok 4.7, meanwhile, is handy for smaller jobs. Reasonably economical, generally useful, and occasionally in need of adult supervision.

Every development team has that one colleague who's enthusiastic, available, and best assigned work with a clearly defined scope.

As for Argon, I'm not ready to judge it.

The pelican was genuinely impressive. The larger task was genuinely disappointing.

Maybe I gave it the wrong kind of assignment. Maybe my use of /goal was poorly suited to the job. Maybe there's a category of work where Argon will absolutely shine.

I've barely spent any time with it. Writing it off now would be ridiculous.

I'll try a few smaller, better-defined tasks after the quota resets.

Preferably ones that don't require a second mortgage in compute credits.

The Part That Isn't Quite As Funny

AI coding tools have made development dramatically faster for me. I'm not about to pretend otherwise.

But somewhere along the way, the job changed.

I spend less time writing every line myself and more time figuring out which model to trust, how much freedom to give it, and when to intervene before things get expensive.

Sometimes an agent will spend an impressive amount of time doing an impressive amount of work.

And somehow leave you with an impressive amount of cleanup.

The frustrating part isn't that it gets things wrong.

It's how diligently it can get things wrong.

Hundreds of lines changed. Logs scrolling endlessly. Checklists full of green ticks.

By the end, you almost feel guilty telling it the result isn't good enough.

Almost.

I'm not particularly interested in which model tops the latest benchmark anymore. Rankings are useful, sure, but they don't review my code, deal with regressions, or finish a task when the weekly allowance runs out.

A model that occasionally performs miracles is exciting.

One that reliably gets me through an ordinary working day is much harder to replace.

Anyway, I'm not upset about the Argon experiment.

Maybe a little poorer in quota, but considerably richer in experience. Which, unfortunately, isn't a currency Google accepts.

I've still got almost a full year to figure out what Argon is good at.

And the pelican?

It's still there, pedaling along with its little red scarf, two fish in the basket, looking like it hasn't got a care in the world.

I envy that bird.

At least it doesn't have to wait for a weekly reset.