Multiplicity

THE NUMBER: HALF. That’s what Anthropic just cut off the price of frontier-grade AI — its own frontier-grade AI. Claude Opus 5 shipped Friday at five dollars per million input tokens and twenty-five out, exactly fifty percent of the Fable 5 flagship the company launched five weeks earlier, and on most of the work that actually pays it beats Fable outright. Nobody undercut Anthropic here. Anthropic undercut Anthropic. Hold that number, because when the price of a capable digital worker falls by half in five weeks, flat, on purpose, from the company that sets the price, it is telling you something the benchmark table is designed to keep you from noticing: the worker was never the scarce thing.

There’s a scene in Multiplicity — the 1996 one, Michael Keaton, back when Keaton would still do a broad comedy — where Doug Kinney, a contractor drowning in a job he loves and a family he’s losing, gets talked into cloning himself. One copy to run the site. Then a second copy, 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 has no business dialing. When Doug asks how this happened, one of the sharper clones delivers the whole thesis of the movie in a shrug: “You know how when you make a copy of a copy, it’s not quite as sharp as the original?”

That is the AI model business in July of 2026. A stable of digital workers, each one cheaper than the last, each one a little different, and somewhere in the lineup a Steve you should not put on the phone with a customer. And on Friday, Anthropic cut the price of a clone in half.

🧬 The Signal: The Company Undercut Itself

Let’s start with what actually happened, because the number is doing more work than the coverage gave it credit for.

Claude Opus 5 shipped Friday, July 24, priced at five dollars per million input tokens and twenty-five per million output, identical to the workhorse Opus 4.8, and exactly half of what Anthropic charges for Fable 5, the model it made its flagship five weeks ago. On the headline Artificial Analysis Intelligence Index, Opus 5 edges Fable by a single point, 61 to 60, which is inside the margin of error and, frankly, noise. The real story is one rung down, on the agentic work that businesses actually pay for. On AA-Briefcase, the benchmark that grades a model on finishing real knowledge work — reports, spreadsheets, presentations built off thousands of input files — Opus 5 beats Fable by 146 Elo points while costing twenty percent less per task. Its cheaper “high effort” setting still beats Fable, at about ten dollars a task against Fable’s twenty-two. On computer use it clears Fable’s best score at roughly a third of the cost. It is Anthropic’s fourth model in two months.

Every outlet filed this as a price war. Wrong movie. This is Clayton Christensen, and Anthropic is running the play on purpose, which is the only way the play ever works. The innovator’s dilemma isn’t that incumbents can’t build the cheaper, disruptive product. It’s that they refuse to, because the cheaper product cannibalizes the fat-margin one they already sell. Kodak had the digital camera in a drawer. Anthropic just did the opposite: took its own highest-margin SKU, the five-week-old flagship, and shot it in the street before OpenAI’s GPT-5.6 Sol or a free Chinese model could. I’ll pick a hand. That’s smart. Disrupting yourself is the single hardest, most valuable move a market leader can make, and almost nobody has the nerve.

And there’s a colder read underneath the brave one, the one the guys on X caught before the press did. The trader Emad Gnia put it flat: shipping Opus 5 while a finished Fable 5.1 reportedly sits on the shelf “isn’t restraint, it’s supply constraint.” Anthropic is famously short on compute; it’s leasing billions in capacity from Meta and SpaceX. You ship the cheaper, lighter model first when you physically cannot serve the expensive one at the volume people want. We wrote on July 20, in I Am Altering the Deal, that Anthropic had cut Fable’s price and dressed it up as a capacity shortage. Five days later they cut the price again, by half, and this time didn’t bother with the costume. Either way you read it — bold self-disruption or forced hand — the tape says the same thing. The price of the best available intelligence is in free fall, and the fall is coming from inside the house.

🖨️ The Copy of a Copy

Here’s where Steve walks in.

These clones do not come off the line identical, and the difference that matters isn’t IQ. It’s temperament. Independent testing pegged Opus 5’s hallucination rate — specifically, how often it answers when it should hedge — at right around fifty percent, up fourteen points from the model it replaces, and higher than the pricier Fable. Sit with that. Anthropic built Opus 5 to be the everyday default, the model now standing at the front of the line on Max and Pro, the one you’ll reach for without thinking. And it is measurably more willing to look you in the eye and make something up. Cheaper labor, more confident, more often wrong. That’s not a bug they’ll patch by Tuesday. It’s a personality.

This is the thread we’ve been pulling all week, and it keeps getting stronger. On July 22, in The Science of Hitting, we argued that intelligence is now cheap and the scarce thing is verification — knowing whether the swing actually connected. Opus 5 is that argument wearing a name tag. The generation got cheaper again on Friday. The grading did not. Worse, the benchmarks crowning Opus 5 were run by Artificial Analysis on Anthropic’s behalf, the company discloses it, and the human-preference boards haven’t even weighed in yet — on LMArena, Fable still leads. So the model that’s “winning” is winning on a scorecard its own maker commissioned, while quietly getting more comfortable bluffing. We told you to grade the grader. This is why.

None of which makes Opus 5 a bad hire. When it’s good, it’s genuinely startling. Handed a Frontier-Bench task where it had to rebuild a machine part in code but was given no way to see the drawing, it wrote its own computer-vision pipeline to pull the geometry from raw pixels, then rebuilt the part. No competing model solved it in five tries. Given a real bug in a package manager, it found the root cause a human community patch had missed. That’s clone number two — the sharp one who can run the whole job site. The point isn’t that the model is dumb. The point is that the same release contains the sharp copy and the Steve, and they answer to the same login, and telling them apart is now your job.

🧑‍🏫 The New Hire Tax

Now the part nobody selling you the upgrade will mention: a new model is not a faster old model. It’s a new employee. And onboarding costs you before it earns you a dime.

The cleanest proof came this week from Every, the media-and-software shop that lives inside these tools and reviews each one honestly. Their verdict on Opus 5: “brilliant in flashes, frustrating in practice.” In its first week the model argued with their instructions, stopped before the work was done, and fought the skills and plugins they’d built up over a year for earlier versions of Claude. It was, by their account, kind of a pain to employ. Then they did something most companies never will. They deleted their own scaffolding — the carefully tuned prompts, the elaborate high-effort settings, the process — and the model got dramatically better. Turned down to low and medium effort, stripped of the harness built for its predecessor, it flew.

Read what that means, because it’s the whole game. The workflows you were proudest of, the accumulated prompt-craft and the plugin stack and the “here’s how we do it here,” those are not assets when the new hire arrives. They’re cultural debt. They were built to manage the last employee’s quirks, and the new one has different quirks, so the scaffolding actively gets in the way. We made this argument on July 19, in Lead, Follow, or Get Out of the Way, about factory electrification — how the productivity gains from electric power took twenty years to show up, because the first generation just bolted a motor onto the old steam-driven line-shaft and ran the plant exactly as before. The gains only came when someone rebuilt the floor around the new power source. Opus 5 is that story compressed into a five-week release cycle. Bolt it onto the harness you built for Fable and it stumbles. Rebuild around it and it soars. Either way, you pay the tuition.

So every upgrade now carries a tax that never shows up on the pricing page: the days your best people spend re-teaching the machine your ways, re-tuning the prompts, discovering which of last month’s tricks now backfire. Cheaper tokens, sure. But the switching cost is human, it’s real, and it recurs every single time a lab ships. Which forces the question the whole issue has been walking toward.

🧰 The One Job Worth the Upgrade

If a new model is a new hire, and hiring costs you real time even when the wage is lower, then the honest question isn’t “is Opus 5 better than Fable.” It’s “which of my jobs is worth onboarding a new worker every five weeks — and which should just keep the one they already trust?”

For almost everything, the answer is: keep the old one. The model you deprecated last month didn’t get dumber. It still writes the deck, clears the inbox, drafts the memo, summarizes the call, exactly as well as it did in June, and it already knows your formats and your voice because you already trained it. Swapping it for a marginally smarter, differently-neurotic replacement buys you a rounding error of quality and costs you a real week of retraining plus a fresh batch of hallucinations you haven’t learned to catch yet. On a cost-per-outcome basis, upgrading that workflow is a negative-return trade. You’re paying tuition to make a job you’d already solved slightly worse for a fortnight.

There is exactly one category that breaks the rule, and it’s the one that’s driven every model release for two years: coding. Code justifies the constant upgrade for two structural reasons nothing else has. First, it’s cheaply and instantly verifiable — the test passes or it doesn’t, the build compiles or it doesn’t, so the new hire’s hallucinations get caught by a machine in seconds instead of shipping to a customer. The Steve problem is contained. Second, the gains compound. A twenty-percent better coding model doesn’t just write today’s function better; it unlocks harder problems that were out of reach last month, and those stack. That’s why the SaaStr crew, running an eight-figure business on AI agents, obsesses over the coding layer while cheerfully leaving everything else alone. It’s also why they made this week’s other quiet, brilliant move: they cut their agent count from thirty back to twenty and watched output go up roughly four times. Not because the models got worse. Because the binding constraint was never model IQ — it was how many different digital employees one human could actually hold in his head. Doug’s problem, exactly. The chaos didn’t come from the clones being dumb. It came from making too many of them and trying to manage them all at once.

So here is the shape of it. Digital labor is getting cheaper by the week, and that is genuinely good news. But cheaper is not free, sameness is a myth, and trust is not included. Each new model is a new hire with its own head and its own tells, and onboarding it burns your scarcest resource, which is the attention of the humans doing the supervising. Spend that attention where the work is verifiable and compounds — your engineering, and precious little else. Everywhere else, resist the shiny new clone. Keep the worker you’ve already trained.

What This Means For You

Downgrade one workflow on purpose this week. Take your heaviest non-engineering task — the deck, the research memo, the customer email — and run it on the model you deprecated last month for a day. If nobody can tell the difference, and they won’t, you’ve been paying frontier prices and frontier onboarding costs for badge-tier work. Standardize the cheaper, older, already-trained model there and save the upgrade budget for the one place it earns out.

Price the onboarding tax before you roll anything new to the team. A new model is a new hire, and Every just showed the bill: rebuilding a year of scaffolding, then discovering it runs better with the scaffolding gone. Before you push Opus 5 company-wide, budget the days your people will spend re-teaching it, and set that number against what the upgrade actually buys. If the honest answer is “a few Elo points on a benchmark Anthropic paid for,” don’t move.

Put a human on the checkbook, and reserve the upgrade for code. Opus 5 makes things up about half the time it’s unsure, worse than its predecessor. Anywhere a confident wrong answer costs real money — finance, legal, anything a customer sees — keep a person on the CC line and a verification step in the loop, the same way you’d supervise a new employee. And concentrate your model-chasing energy on the engineering workflows where a machine can catch the mistakes for free and the gains stack. That’s the only room where hiring the newest clone every five weeks actually pays.

Three Questions We Think You Should Be Asking Yourself

1. Which of my workflows are verifiable, and which just feel productive? The verifiable ones — where “done” and “correct” can be checked cheaply — are the only place the constant upgrade earns its onboarding cost. Everywhere else you’re paying tuition to re-solve a problem you’d already solved. If you can’t name which of your jobs has a real oracle for “is this right,” you can’t know where to spend the upgrade budget.
2. What is my onboarding tax, in days, the next time a lab ships? If you don’t know how long it takes your team to re-tune prompts, rebuild plugins, and relearn a new model’s tells, you are treating “upgrade” as free when it is one of the most expensive recurring costs in your AI budget. The shops that win the next two years are the ones that measure it and refuse the trade when the math says no.
3. How many digital employees can one of my humans actually hold in their head? SaaStr found the number was smaller than the demo promised, cut their agent count by a third, and quadrupled output. Doug found it out the hard way with a house full of Steves. Every clone you add costs attention, and attention is the resource that doesn’t get cheaper when tokens do.

“You know how when you make a copy of a copy, it’s not quite as sharp as the original?”
— Steve, Multiplicity (1996)

— Harry and Anthony

Signal/Noise by CO/AI is published most weeknights from New Canaan, Connecticut. The point is to make you the smartest person in the room without taking more than fifteen minutes of your morning. If we did, forward it to one person. If we didn’t, hit reply and tell us why.

---

Sources

• Introducing Claude Opus 5 — Anthropic, Jul 24, 2026 ($5/$25 per M tokens, same as Opus 4.8; new default on Max, strongest on Pro; five effort settings; the FreeCAD computer-vision and package-manager root-cause examples; cyber classifiers fall back to Opus 4.8, ~85% less often than Fable)
• Anthropic Ships Opus 5: Half the Price of Fable 5 — Trending Topics, Jul 26, 2026 (Intelligence Index 61 vs Fable 60; hallucination-when-uncertain ~50%, up 14 pts, worse than Fable; Fable 5 = $10/$50; effort dial spans 8x tokens / 407 Elo; benchmarks Anthropic-commissioned; LMArena not yet rated, Fable leads human preference)
• Claude Opus 5 is the new leader on AA-Briefcase — Artificial Analysis, Jul 24, 2026 (+146 Elo over Fable at 20% lower cost per task, $17.79 vs $22.30; high-effort beats Fable at $10.41; GDPval-AA v2 +114; OSWorld beats Fable’s best at ~⅓ cost)
• Vibe Check: Claude Opus 5 Is Brilliant in Flashes, Frustrating in Practice — Every (Dan Shipper, Katie Parrott), Jul 24, 2026 (fought the scaffolding built for older Claude models; got dramatically better once instructions were deleted; low/medium effort stronger for coding)
• We Peaked at 30 AI Agents. Now We’re Coming Back Down to 20 — SaaStr (Jason Lemkin), Jul 24, 2026 (30→20 agents, ~4x output; “optimize for how many agents a human can hold in their head”)
• @emadgnia on X, Jul 26, 2026 (via Aligned News) — “shipping Opus 5 before a finished Fable 5.1 isn’t restraint, it’s supply constraint”
• CO/AI prior issues this builds on: I Am Altering the Deal (Jul 20 — Anthropic’s price cut in a capacity costume; the one-point moat); The Science of Hitting (Jul 22 — verification is the scarce resource; grade the grader); Lead, Follow, or Get Out of the Way (Jul 19 — the missing management layer and the factory-electrification lag)
• Multiplicity (1996) — Doug Kinney, the clones, and Steve, the copy of a copy