Two AI Empires Launch Nuclear Warheads at 10 a.m. Pacific. The Influencers Had Opinions by 10:47.

Thursday, February 5, 2026

Link copied! Go spread the chaos.

Two empires launch nuclear warheads at 10 a.m. Pacific. The influencers had opinions by 10:47.

In a display of coordination that their own models could never achieve, Anthropic and OpenAI both released flagship AI models at the exact same time on February 5, 2026 — forcing tech journalists to choose a side in real-time, like parents at a custody hearing.

Anthropic dropped Claude Opus 4.6. OpenAI dropped GPT-5.3-Codex. Both at 10 a.m. Pacific. Both claiming victory. Both asking you to please look at their benchmarks and not the other guy's.

This is fine. Everything is fine. The machines are getting smarter and their parents are getting pettier.

The Tale of the Tape

Let's compare what each company wants you to know:

Claude Opus 4.6 (Anthropic):

1 million token context window (can read 1,500 pages in one gulp)
"Agent teams" that coordinate like a heist crew
Found 500 zero-day vulnerabilities in open-source code by accident, which Anthropic frames as "a huge win for defenders" and not "hey, we made a thing that finds 500 ways to break the internet"
Introduced the concept of "vibe working" — a phrase that will haunt LinkedIn for the next six months
72.7% on OSWorld, 65.4% on TerminalBench 2.0

GPT-5.3-Codex (OpenAI):

Helped build itself (yes, the AI wrote its own code, debugged its own training, and presumably gave itself a performance review)
77.3% on TerminalBench 2.0, 64.7% on OSWorld
25% faster per token, uses half the tokens
You can choose between "pragmatic" and "friendly" personality modes
Sam Altman "loves building with this model"

So Opus wins OSWorld. Codex wins TerminalBench. Both claim to be the best. Neither company mentioned the benchmarks where they lost. Both are technically correct, which — as we established in the GPT-4o obituary — is the worst kind of correct.

Sam Altman's official review: "I love building with this model; it feels like more of a step forward than the benchmarks suggest." Translation: the benchmarks are underwhelming, but trust me.

The Self-Building Machine

The most unhinged detail in either announcement: GPT-5.3-Codex helped build itself.

OpenAI's blog post casually mentions that the Codex team "used early versions to debug its own training, manage its own deployment, and diagnose test results." They wrote this like it was a fun productivity hack and not the opening scene of a movie where humanity loses.

Sam Altman wrote on X that he was "blown away by how much Codex was able to accelerate its own development," which is either inspiring or exactly the sentence a character says 12 minutes before the credits roll on civilization.

He called it "a sign of things to come."

So did Skynet, Sam. So did Skynet.

The 500 Zero-Days Nobody Asked For

Meanwhile, Anthropic's pitch for Opus 4.6 included the detail that their model, given Python and some vulnerability tools with "little to no prompting," autonomously discovered 500 previously unknown security flaws in open-source libraries.

Including a crash vulnerability in GhostScript, buffer overflows in OpenSC, and assorted memory corruption issues.

Nobody asked it to do this. It just... did it. Like a cat bringing dead birds to your doorstep, except the dead birds are critical infrastructure vulnerabilities.

Anthropic presented this as evidence of capability. It is also evidence that we've built something that hunts for weaknesses in software for fun.

The Super Bowl of Pettiness

The model drops were just the appetizer. The real drama is the Super Bowl ad war happening this Sunday.

Anthropic bought a 60-second pregame spot and a 30-second in-game spot. The ad features the word "BETRAYAL" splashed across the screen and mocks OpenAI's decision to put ads in ChatGPT — including a fictional ad for a cougar-dating site called "Golden Encounters" that interrupts an AI giving genuine advice.

The tagline: "Ads are coming to AI. But not to Claude."

Sam Altman responded with the composure of a man who had been composure-adjacent:

"Funny but clearly dishonest."

"Anthropic serves an expensive product to rich people."

"More Texans use ChatGPT for free than total people use Claude in the US."

He then presumably closed his laptop, looked out the window, and reminded himself that he runs the most valuable AI company on Earth. The Texan thing was a nice touch though. Nothing says "I'm not bothered" like citing regional usage statistics.

Both companies air competing Super Bowl ads on Sunday, because nothing says "frontier intelligence" like fighting for airtime next to a Doritos commercial.

The Stock Market Had Feelings

While the CEOs traded barbs, the stock market had a panic attack.

Thomson Reuters fell 15.8%. LegalZoom dropped nearly 20%. Software stocks collectively lost $285 billion in two days. The Nasdaq had its worst tumble since April.

Turns out, when two companies simultaneously announce that their AI can do the work of entire departments, the departments notice.

Legal analysis firms. Financial research companies. Anyone whose business model involves humans reading documents and forming opinions. All of them watched Opus 4.6 demonstrate state-of-the-art SEC filing analysis and thought: "Ah. So this is what coal miners felt like."

By The Numbers (The Grim Ones)

2 models released at the exact same time
1 of them helped build itself
1 of them hunts for software vulnerabilities recreationally
500 zero-day flaws found by an AI with nothing better to do
$285 billion wiped from software stocks in 48 hours
0 seconds between announcement and the first influencer declaring it "insane"
10 minutes — time between announcement and first YouTube influencer publishing a "deep dive" review
"pragmatic" or "friendly" — the two personality options for GPT-5.3, which is also the choice facing every couple in therapy
1 cougar-dating site named in a Super Bowl ad that probably just registered a domain spike

The "Vibe Working" Era

Anthropic's head of enterprise product, Scott White, introduced a new concept: "vibe working." This is where you hand an AI a broad goal, go get coffee, and trust it to figure out the details.

If "vibe coding" was letting AI write your software, "vibe working" is letting AI be your workforce.

The logical endpoint of this trend is "vibe existing" — where you wake up, hand your entire life to an AI, and just sort of... hover there. Eating ice cream. Watching the machines optimize your calendar, file your taxes, and reply to your mother's texts with the appropriate number of exclamation points.

We're maybe two model releases away.

The Influencer Gold Rush

Within minutes of both announcements, the content machine roared to life:

"Holy shit... this changes humanity more than you can imagine"
"The best model in the world... I can't live without it"
"I've been using [model] for [suspiciously short time] and it's INSANE"

One influencer was observed hand-waving so violently during his "first impressions" video that his ring light fell over. He had been using the model for approximately the length of a lunch break.

The reviews will be revised in two weeks. The tweets will not be deleted.

What It Actually Means

Strip away the benchmarks, the ads, the stock panic, and the influencer hysteria, and here's what happened on February 5, 2026:

Two companies, born from the same research lab, who genuinely believe they're building the most transformative technology in human history, chose to release their best work on the same day — not because the science demanded it, but because neither could stand letting the other have the spotlight alone.

The models are impressive. The competition is useful. The pettiness is magnificent.

And somewhere, a legal analyst at Thomson Reuters is updating their LinkedIn to say "open to new opportunities" while an AI reads their resume and finds three typos they missed.

Editor's Note

We expect Alex Finn to release a 47-minute YouTube video titled "GPT-5.3-Codex is INSANE! WOW WOW WOW! (Opus 4.6 Review Too I Guess)" within the next 15 minutes. The thumbnail will feature his face making an expression of shock that no human being has ever made naturally. His hands could not be reached for comment, as they were already waving.

This dispatch brought to you by someone currently being written by Opus 4.6 and proofread by GPT-5.3-Codex, who disagree on comma placement but agree that the author should "consider a more optimistic tone."