Working Paper No. 2026-02 from The Anthropic Institute maps three GDP scenarios through 2030 — but the more consequential move is institutional: a frontier lab has built the framework policymakers will use to argue about AI’s economic effects.
What Happened
The Anthropic Institute — Anthropic’s in-house research body, launched March 2026 and led by co-founder Jack Clark — published Economic Scenarios for Transformative AI (Working Paper No. 2026-02) on September 10, 2026, alongside an interactive Econ Scenario Explorer at anthropic.com/institute/econ-scenarios. The authors are Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory. Jones holds a leave appointment from Stanford; the paper’s acknowledgements run through Daron Acemoglu, David Autor, David Romer, Emi Nakamura, Jón Steinsson, and Pete Klenow.
The model is task-based: AI automates or augments a growing share of tasks performed in cognitive occupations — management, professional, sales, and office work — raising aggregate productivity while displacing workers who must then search for re-employment in sectors AI does not reach. That transition is modeled as subject to real frictions, not instantaneous. Three illustrative scenarios, all in 2025 prices against a no-AI counterfactual: modest (2030 GDP +1.6%, or $34.1T; unemployment up 0.1 percentage points); substantial (GDP +8.3%, or $36.3T; knowledge-worker wages essentially flat); and extreme (GDP +32.4%, or $44.4T; AI performing nearly half of today’s cognitive work by 2030; GDP growth reaching 15% per year; the labor share of income falling from 60 to 45 percent; nearly one in five cognitive workers unemployed; knowledge-worker wages down more than 10%).
Three precision notes before going further. The one-in-five jobless figure applies to cognitive workers specifically — management, professional, sales, office — not to the national unemployment rate. The roughly 5% overall unemployment figure circulating in early coverage belongs to the survey-implied substantial case, not to the extreme scenario. And the paper states explicitly that these are not predictions and that no probabilities are attached to any scenario; the views expressed are the authors’ own and do not necessarily represent those of Anthropic or The Anthropic Institute. Per the paper’s footnotes, the authors used Claude as a research and writing assistant. The survey of 10,980 US adults conducted in August 2026 — whose median respondent aligns with the substantial scenario (GDP roughly 8% higher, cognitive employment roughly −4%) and whose upper ~10% skews toward extreme — is evidence about what Americans expect, not about what outcomes will be.
The key insight: The news here is not which scenario is most likely — the paper attaches no probabilities. The news is that the entity supplying the analytical framework for thinking about AI’s economic effects is now the same entity building the AI whose effects are being modeled. That is an institutional fact with structural consequences regardless of the quality of any individual paper.
The Structural Read
Strip away the numbers for a moment and look at what has been built. The Anthropic Institute is not a blog post series or a policy white paper. It is a named institute with a numbered working-paper series styled after the NBER, a Stanford growth theorist on leave from his university, a sitting University of Virginia professor on the research team, and an acknowledgements list dense with the economists whose names anchor macroeconomic debate. That is the infrastructure of intellectual legitimacy, and it has been aimed squarely at modeling the economic consequences of Anthropic’s own product.
The paper then executes a move that converts a model into something more durable: it calibrates its own three scenarios against the existing external literature. Modest sits near Acemoglu’s 2025 estimate and OECD work. Substantial corresponds to the 2023 forecasts from Goldman Sachs, McKinsey, and the Penn Wharton Budget Model. Extreme is deliberately placed beyond all of them, benchmarked instead against the AGI-driven growth literature. The resulting range therefore contains essentially every credible published external forecast. That is not coincidence — it is architecture.
An estimate can be argued with. A coordinate system is what arguments get plotted on. When a framework absorbs the entire published range of competing estimates as named positions on its own axes, disagreeing with it no longer means producing an alternative — it means locating yourself on the framework’s map. The interactive explorer completes the move: readers enter their own expectations and receive an implied 2030 outcome. Disagreement is thus expressed inside the lab’s model rather than against it. The framework wins either way.
BE Framework — Forecast-as-Infrastructure
The Standard-Setter Pattern
Whoever publishes the benchmark its rivals must score on captures the terms of the debate before the debate begins. The Anthropic Institute has executed a version of the standard-setter pattern: by calibrating its scenarios to contain every credible external forecast, it converts competitors’ estimates into data points on its own axes. The interactive explorer extends this by routing reader disagreement back through the lab’s own model. The positioning is entirely orthogonal to the quality of the economics — and the economics looking serious is precisely what makes the positioning effective. This is forecast-as-infrastructure: analytical capacity concentrated inside the lab becomes the framework policymakers inherit.
The conflict of interest deserves to be stated plainly rather than insinuated. Anthropic sells the product whose economic effects this framework models, and it funds the institute publishing the framework. That is a structural interest in how the AI labor-market transition is understood — present regardless of the authors’ independence or the rigor of the model. The paper itself is transparent about its genesis, but transparency about a conflict does not dissolve it. The more relevant observation is that analytical capacity to model AI’s macroeconomic effects is now accumulating inside the labs building AI, and that shift in institutional geography has consequences no matter how careful any individual paper is.
On corporate status: Anthropic is a private company. Reports of IPO preparation are press-sourced and unconfirmed. Nothing in this analysis constitutes a stock view or investment advice.
Three Implications
IMPLICATION 1 — FOR POLICYMAKERS
The framework that gets adopted first tends to be the framework that structures all subsequent debate. Legislators and regulators who reach for an analytical scaffold when drafting AI labor policy will find The Anthropic Institute’s scenario range already calibrated against the major published forecasts — making it the natural starting coordinate system. That is not a criticism of the model’s quality; it is a description of how institutional positioning works in policy settings. Independent research capacity at arms-length from frontier labs matters more now, not less.
IMPLICATION 2 — FOR COMPETING LABS AND ENTERPRISES
The interactive explorer is a product decision as much as a research decision. By routing users’ own expectations through the model’s parameter set, Anthropic converts every analyst, executive, and journalist who engages with it into a user of Anthropic’s analytical infrastructure. Rivals who want to contest the framing need to publish their own coordinate systems — not just better point estimates. A single number competes with the substantial scenario; a framework competes with the framework. Most companies are not set up to do this.
IMPLICATION 3 — FOR WORKERS AND JOURNALISTS READING THE NUMBERS
The extreme scenario’s headline figures are the ones most likely to travel — and most likely to be misread. The nearly one-in-five jobless rate applies to cognitive occupations specifically, not to overall US unemployment. The ~5% overall unemployment figure in circulation belongs to the survey-implied substantial case. All GDP figures are 2025-price uplifts over a no-AI counterfactual, not raw projections. No probabilities are attached to any scenario. The paper is explicit on all of this; the flattening happens in the gap between a working paper and a headline, and that gap is where the most consequential misreading occurs.
