Anthropic's own model pays capital 81 per cent more in the branch where wages stand still

Divide each scenario's output gain by its task share and the model's shape appears. Four per cent of tasks buys 1.6 per cent of extra GDP, which is 0.4 of a percentage point of output per point of task share ^[3][1]; twelve per cent of tasks buys 8.3 per cent, or 0.69 ^[4][2]; close to a third of tasks buys 32 per cent, which if you read a third as 33.3 comes to 0.96 ^[5][3]. The last point of adoption is worth about two and a half times the first ^[7]. Returns to diffusion accelerate inside this framework, which is why the middle branch is the awkward one to underwrite: it comes out to a quarter of the extreme case, not half.

The split is the figure worth an allocator's attention, more than the output dial. Anthropic's extreme branch takes the labour share of national income from 60 per cent to 45 per cent ^[6], and 1.32 multiplied by 45/60 is 0.99, so total wages finish where they started even though the economy is a third larger, which is what the company's own write-up says ^[5][6]. Invert it. Capital's share goes from 40 to 55, and 0.55 times 1.32 over 0.40 is 1.81, so capital income rises roughly 81 per cent against a wage bill that does not move ^[4]. The substantial branch publishes no labour share, so that same sum cannot be run at 12 per cent of tasks; the material gives unemployment of about 4.6 per cent and a 4 per cent decline in knowledge-sector jobs and stops there ^[4].

The scenario numbers reach us via Axios, and the tool was built by Anthropic's own economics team on a framework that sorts the economy into cognitive occupations and everything else ^[3][2]. That authorship does not make the conclusion wrong, but it does tell you which output to interrogate. Any model in which software substitutes for cognitive tasks will hand the gain to whoever owns the software; income moves toward capital in all three branches, which it does ^[7], and the interesting question is how far and how fast, entirely a function of the elasticity the user dials in. Jack Clark, who told NPR he sees no slowdown in the pace of development but expects the path into the economy to be slower than most observers assume ^[11], is describing a slower march up the same curve, not a different curve.

What would break this read is the labour share holding near 60 per cent at high task adoption, because that would mean AI is complementing cognitive work rather than replacing it, and the 32 per cent of extra output would land in pay packets rather than in margins. The technical paper by Anton Korinek, Charles Jones, Szymon Sacher, Tess Cotter and Peter McCrory is where that elasticity lives, and the authors are explicit that none of the three paths carries a probability ^[8]. Nearly 11,000 US adults surveyed by Anthropic put the median expectation at about 10 per cent more GDP and roughly 5 per cent unemployment, 1.7 points of output above the substantial scenario ^[10][6]. The public, in other words, has already picked a branch that the model itself has not.