Anthropic chief executive Dario Amodei has published a plan urging government-backed coordination to pace frontier AI development, framed around safety checkpoints and independent reviewers with publication rights. A September 13 essay on Cogito Ergo Sum argues the plan also happens to serve Anthropic’s balance sheet, and that the two motives are worth separating rather than taking on faith.
The essay’s core claim is straightforward: slowing the arrival of new models lets existing ones keep earning at premium prices for longer. Price erosion in frontier AI is fast. Citing Artificial Analysis pricing data compiled by CatalystNeuro, the essay finds that the cost of buying a given level of benchmark performance has halved roughly every 46 days across five tracked models, each with a price history of 90 days or more. A GPT-6 Astra configuration in the September 13 Pareto-frontier snapshot scored higher than a Fable 5 configuration while costing $1.72 per task against $8.75, illustrating how quickly a costlier model can lose its edge to a cheaper one at similar capability.
That dynamic creates a structural trap the essay lays out in three stages: a lab invests to build a capability lead, rivals compete that lead away as they catch up, and the lab invests again to restore its premium. Each cycle can leave a wider gap between development spending and the revenue available to repay it. The essay’s own financial evidence shows how wide that gap already runs: OpenAI’s reported 2025 gross profit of $5.57 billion covered only about 29 percent of its $19.18 billion research and development expense, even as gross margin rose from roughly 28 percent in 2024 to 43 percent in 2025.
Coordinated pacing, the essay argues, interrupts that cycle in a way no single lab can manage alone. A lab that unilaterally slows down risks ceding the frontier to rivals who do not, so binding restrictions on Chinese compute access and unauthorized distillation, both features of Amodei’s proposal, make slowing down “game-theoretically feasible” only when competitors face the same limits.
The essay does not dispute that safety concerns inside frontier labs are sincere. It cites Amodei’s own argument that a slower pace leaves room to investigate dangerous behavior and test models properly, and notes a separate AI Futures proposal that would reserve just 5 percent of compute for capabilities research after allocating 70 percent to serving customers and 25 percent to safety work. Its point is narrower: sincerity does not make a lab a disinterested party in writing the rules that govern its own industry, and being profitable does not make a lab any less eager to shape those rules. The piece notes Anthropic’s own reported $47 billion revenue run rate as of May, and what it describes as a Reuters account, not independently verified, that the company posted positive adjusted operating income for a second quarter running.
For operators evaluating AI vendor roadmaps, the essay’s framing is a useful filter rather than a verdict: any lab lobbying for development limits is also lobbying to extend the shelf life of the model it already sold you. Anyone building procurement or competitive strategy around expected pricing should treat “price of intelligence halves every 46 days” as the base case to plan against, not the exception that regulation might interrupt.
Based on an analysis essay published September 13, 2026 on Cogito Ergo Sum (cogito-ergo-sum.dev).
