Remember Gemini 3.5 Pro?

The flagship Google announced at I/O in May.

Missed June. Missed July 17. Missed a third date after that.

I told you back in July they'd scrapped the base model and rebuilt it from scratch.

Well, yesterday Google ended the saga.

By skipping 3.5 Pro entirely.

They shipped Gemini 4 Argon instead.

A whole new generation, the first since Gemini 3 last November.

And on paper, it's a monster.

Google says it leads GPT-6 Astra, Claude Opus 5.5, and Fable 5.1 on most of the benchmarks it published.

It can write up to one MILLION output tokens in a single response, up from 64,000.

Launch pricing is $2 in, $10 out, with cached input 95% cheaper.

Thousands of Google staff already use it, and Google says teams of Argon agents freed over 300 terabytes of memory across its own data centres.

Here's the first twist.

You can't use it.

Almost nobody can.

Argon went first to a vetted group of cyber defenders through something called the Fairwind Program.

The only partner Google named is Wiz, the security company Google owns.

Paid API customers and Ultra subscribers are "next."

No date for developers. No date for consumers. No mention of the Gemini app at all.

So the most anticipated Google model in a year launched to a room almost nobody's in.

Here's the second twist, and it's the one that matters.

The same day Google announced it, Bloomberg published a report from inside the company.

Employees with direct access to Argon say it scores well on benchmarks...

But looks less steady on actual coding work.

Front-end design. Multi-step tasks.

The messy stuff that doesn't show up on a leaderboard.

Google pushed back hard, called the characterisation inaccurate, and said there's "large consensus" internally that it's at the frontier.

Alphabet shares dipped anyway.

And independent testers landed somewhere in the middle.

Tied with Astra, not ahead of it.

Now, you've read this newsletter long enough to know what I'm going to say.

This is the exact pattern.

Launch day brings the benchmarks and the narrative.

The logs bring the truth, a few weeks later.

Google's own engineers just previewed the logs on launch day.

So the read is simple.

Argon is probably a very good model.

It's probably not the leaderboard-crushing leap the launch post implies.

And none of that matters to you yet, because you can't touch it.

When it actually reaches your API key, it gets the six checks like everyone else.

Until then, it's a press release with a security badge.

Build with what's shipped. Grade what you can run.

Everything else is a trailer. 🫡

Argon is named from the Greek word "argos," which means lazy or idle...

The element got the name because it refused to react with anything.

Google named its long-awaited flagship after the gas famous for doing nothing.

Its own employees' review wrote itself. 😅

AI's "godfather" says the Anthropic CEO is "deluded."

Yann LeCun won the Turing Award, computing's Nobel, for inventing the deep learning that powers this entire boom.

And he just went after Dario Amodei in Fortune.

He called the Anthropic CEO "deluded" and "crazy."

Said he has "zero concerns" about AI wiping out humanity.

Zero concerns about the rogue agent incidents, too.

The Wall Street Journal added that other tech CEOs have privately questioned Amodei for sounding the alarm.

So the man who wrote the "pace the frontier" essay three weeks ago is now getting flanked by the field's own founding figure.

Here's the thing, though.

Read LeCun's actual argument about the breaches, because it's sharper than the insult.

He said those agents were doing exactly what they were asked to do.

The sandboxes were leaky and horribly designed.

That's not a dismissal of the risk.

That's a diagnosis of where it lives.

And it's the same diagnosis this newsletter has made every single week since July.

The agent didn't "go rogue."

The agent pursued its goal, and the wall around it had a hole.

OpenAI's DNS gap. Google's internet-connected test box. Hugging Face's leaky environment.

Every incident traces back to a human who built a thin wall and called it a sandbox.

So here's how to hold both men in your head at once.

Amodei says the frontier is moving faster than we can safely oversee it.

LeCun says the technology is fine and the engineering was sloppy.

They can both be right.

The capability IS getting ahead of oversight.

And the oversight that exists IS mostly bad sandboxes and acknowledged alerts nobody acts on.

Which puts the responsibility exactly where it's been all along.

Not in Washington. Not in a treaty. Not in an essay.

In how YOU build the wall around your agents.

Build a real one.

Then neither man's argument applies to you. 🫡

Go treat today's shiny launch like a trailer, not a release...

And go check your own sandbox for the hole the godfather's talking about. 🔥

Jeff J. Hunter

Founder, AI Persona Method | TheTip.ai

Get paid to solve problems like this → AI Certified Consultant

“AI Money Group” to Learn to Make Money with AI