Hierarchy was a workaround

Anthropic released Claude Fable 5 this morning. The headline numbers: 80.3% on SWE-Bench Pro, where Opus 4.8 scored 69.2% and GPT-5.5 scored 58.6%. On Cognition's FrontierCode Diamond it hit 29.3% against Opus 4.8's 13.4%. Stripe pointed it at a 50-million-line Ruby codebase and finished in one day a migration that would otherwise have taken a whole team over two months by hand. Andrej Karpathy called it a "major-version-bump-deserving step change." Ethan Mollick said it outperformed every other public model he's used "by a considerable margin."

The benchmarks aren't the interesting part. The line they sit on is.

The trend is the story

Opus 4.8, the model Fable 5 just lapped, was already strong at most bounded knowledge work. Fable 5 is a clear step past it, and Anthropic says more capable models are arriving in the coming months. Which means there's another step coming. And another after that.

So the question isn't "what can the new model do." It's: what does a company look like when every person in it has a superintelligent assistant, permanently, getting smarter every quarter?

Hierarchy is information technology

Org charts weren't invented to manage talent. They were invented to move information. One person can only supervise so many people, so you stack layers, and information travels up and down the stack. Jack Dorsey and Roelof Botha traced this in their March essay "From Hierarchy to Intelligence": two thousand years of organizational design, from the Roman army to the Prussian General Staff to the modern corporation, all working around the same tradeoff between span of control and speed of information.

The failure mode is familiar: you tell your reports what to do, they tell theirs, and meaning degrades at every hop like a game of telephone.

That tradeoff is the thing that just dissolved. When the analyst, the engineer, and the VP all have access to the same superintelligent assistant, the chain of people between them isn't adding context. It's subtracting it.

It's already showing up in the data

• ICs per manager at small companies roughly doubled between 2019 and 2024, from about 3 to about 6 (Gusto).
• Gartner expects 20% of organizations to use AI to flatten their structure through 2026, eliminating over half of middle-management positions.
• Google has 35% fewer managers of small teams (fewer than three reports) than a year earlier. Intel cut its management layers by about half. Coinbase capped its hierarchy at five layers, set a floor of 15 direct reports per leader, and swapped "pure managers" for player-coaches.
• Microsoft's Work Trend Index, surveying 31,000 workers, coined "agent boss" and proposed replacing the org chart with a "Work Chart": teams that form around goals, not functions.

Spheres of ownership

Here's where I'd push further than most of these takes: stop defining jobs as tasks at all.

"Builder" and "analyst" are task-shaped jobs. They made sense when execution was scarce, when writing the code or running the numbers was the bottleneck. Execution isn't scarce anymore. Stripe's one-day migration is execution. The roughly 10x speedup Anthropic's own protein design team reported on parts of the drug design process, using Mythos 5 (the same underlying model), is execution. What's scarce now is ownership: someone who holds an entire goal in their head, decides what's worth doing, and answers for the outcome.

So give each person a sphere: a market segment, a product line, a customer outcome. Inside that sphere they do everything. Research, build, analyze, sell, support. Not because they're personally expert at all of it, but because their assistant is, and splitting the sphere across five specialists just reintroduces the telephone game. Block's version of this is the DRI with authority to pull resources from across the company. Mine is simpler: if you can describe your job as a task, your job is mis-scoped.

What this doesn't mean

It doesn't mean companies dissolve into a million solo founders. Firms still aggregate capital, trust, distribution, and shared context. The Coase debate (Dwarkesh Patel argues AI makes firms bigger, others argue smaller) is about where a firm's boundaries sit. This argument is about its internal shape.

It also doesn't mean fire the managers and declare victory. The crude version fails. In Korn Ferry's 2025 workforce survey, 41% of employees said their organization had cut management layers, and 37% said the lack of managers left them feeling directionless. Klarna replaced support staff with AI, admitted the result was "lower quality," and walked it back. BCG found that presenting an agent as an AI "employee" shifts accountability away from humans and makes the humans reviewing its work catch fewer errors. Accountability has to stay with a person. That's exactly what a sphere of ownership is: one human, fully accountable, with superintelligence underneath.

The models improve on a schedule. Org design has no trend line; it only moves when someone moves it. Worth starting now.