SIGNAL / NOISE

A Copy of a Copy

In Multiplicity, Michael Keaton plays a guy so underwater at work and at home that he clones himself. One copy to run the site. Then a second, because one wasn't enough. Then the copies, unsupervised, make a copy of their own, and out comes clone number four, Steve, who has the approximate cognitive wattage of a Xerox machine left running on the same page all night. Steve is sweet. Steve is useless. Steve is the one who eats the wrong sandwich and dials a woman he shouldn't. "You know how when you make a copy of a copy," one of the sharper clones shrugs, "it's not quite as sharp." Cheap labor. Iffy judgment.

That's the AI model business now, and on Friday Anthropic cut the price of a clone in half.

Claude Opus 5 shipped at five dollars per million tokens in, twenty-five out, exactly half what Anthropic charges for Fable 5, the flagship it launched five weeks ago. Opus 5 noses past Fable by a point on the headline intelligence index and beats it clean on the agentic work that actually pays: knowledge tasks, computer use, coding, at a third to a half the cost. Read that twice. The company undercut its own best product by fifty percent, five weeks out, and did it to itself before anyone else got the chance.

Every outlet is filing this as a price war. Wrong frame. This is Clayton Christensen, run on purpose. Anthropic is eating its own crown jewel before OpenAI or a free Chinese model eats it first, and disrupting yourself is the only version of the innovator's dilemma anybody actually wins. Smart. Full stop.

But here's the copy-of-a-copy part. These clones don't come off the line identical. Independent testing put Opus 5's hallucination rate, when it's unsure, at fifty percent, up fourteen points, and worse than the pricier Fable. So the cheaper worker you're about to make the default is also the one more likely to look you dead in the eye and make something up. Same rough IQ as the last hire, different personality, and a real question about which one you hand the checkbook. Clone number four could frame a house. You just didn't want him talking to clients.

And every one of these hires costs you before it earns a dime, because a new model is not a faster old model. It's a new employee. Every found that out this week: Opus 5 fought the workflows and plugins they'd built for the last Claude, argued with instructions, quit early. Then they deleted all of it, and it got dramatically better. The scaffolding you were so proud of is the onboarding tax on the next hire.

Which lands the question nobody selling you the upgrade wants asked: what job actually needs a smarter worker every five weeks? Coding, mostly, because code compiles or it doesn't, the win is checkable, and it compounds. For the deck, the memo, the inbox, the model you deprecated last month was already good enough and it already knows how you like things. Stop interviewing. The best employee isn't the newest one. It's the one you've already trained.

At COAI today: the full Signal/Noise — why exactly one job in your building justifies the constant upgrade, and how to price the onboarding tax on all the others — is live at getcoai.com.

Which of your workflows actually need the newest model, and which just need the one they've already learned. If Opus 5 has you reaching for the upgrade, that's the conversation we want to have.

Outsider Labs Advisory →

ONE — A NUMBER THAT SUMMARIZES THE DAY

HALF. That's what Anthropic just cut off the price of frontier-grade AI — its own. Claude Opus 5 shipped Friday at five dollars in, twenty-five out per million tokens, exactly fifty percent of the Fable 5 flagship it launched five weeks earlier, and it beats Fable on most of the work that pays. Nobody undercut Anthropic. Anthropic undercut Anthropic. When the price of a capable digital worker falls by half in five weeks, the worker was never the scarce thing.

THREE — ACTIONS TO TAKE TODAY

Downgrade one workflow on purpose today. Most non-coding work — decks, memos, first-draft analysis, inbox — ran fine on the model you deprecated last month. Take your heaviest non-engineering task, run it on last quarter's model for a day, and see if anyone can tell. If they can't, you've been paying frontier rates for badge work.

Price the onboarding tax before you roll out Opus 5. A new model is a new hire. Every rebuilt its whole Claude setup to make Opus 5 behave, then got better results after tearing the scaffolding out. Before you push a new model to the team, budget the days it'll take to retrain your prompts, plugins, and habits, and set that against whatever the upgrade actually buys.

Put a human on the checkbook jobs. Opus 5 makes things up roughly half the time it's unsure, worse than the model it replaces. Anywhere a confident-but-wrong answer costs real money — finance, legal, anything customer-facing — keep a person on the CC line and a verification step in the loop. Same trust rule you'd use on any new employee.

FIVE — STORIES TO KEEP YOU INFORMED

Monday, July 27

  • Anthropic undercut itself by half and called it a Friday. (Full analysis above.) Opus 5 ships at half Fable 5's price and beats it on the agentic work that pays. The self-disruption is the signal. The benchmark leaderboard is the noise.

  • Washington's open-weight brawl puts Nvidia against Anthropic. Twenty-plus companies — Nvidia, Microsoft, Meta, IBM, Hugging Face — told the government not to lock down open models, leaving OpenAI and Anthropic isolated as the rent-seekers. Kimi K3's weights drop today. Follow the incentive, not the flag.

  • SaaStr cut its AI agents from 30 to 20 and output went up 4x. The bottleneck was never model IQ. It's how many agents one human can hold in their head. Everyone selling you "deploy a hundred agents" has the arrow pointing the wrong way.

  • Kalanick's back, with $1.7 billion for atoms. The Uber founder raised $1.7B for Atoms, an industrial-AI and robotics play, the same week defense tech crossed $12.3B for the year. As model margins compress, the premium is rotating to the things you can't distill overnight.

  • Ninety cents of every AI venture dollar now comes from a corporate arm. Per PitchBook, corporate VCs are ~90% of 2026 AI funding, Nvidia alone $189B this year. When the same balance sheets are investor, supplier, and customer, "record funding" and "healthy market" stop being the same sentence.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate

Keep Reading

View more
caret-right