Questions and Boundaries

Primary questions

What did Altman fix? A broad capability level, illustrated by GPT-4 and GPT-4o; the numerical example is token price, not job cost. [1]
What did Epoch fix? A score threshold on each named benchmark, then the cheapest estimated model/settings bill to reach it on each date. [2]
Is 13× a forecast? No. It is an annualized summary of about three years of observed/fitted cost-frontier declines. Continuation is uncertain. [2]
Does the 725× GPQA pair set that rate? No. It is a single threshold and exceptionally steep interval; the reported average pools five benchmarks and many score levels. [2, 4]
Why can the meters diverge? Reasoning tokens per question, performance of each token, model mix, and optimization over settings all change the cost of an answer beyond a token’s sticker price. [2]
How robust are the observations? The 75% science-example rows are in Figure 2; Epoch cautions that model coverage and prices are noisy, and tasks can be optimized for benchmark questions. [2, 4]
What would change the decision to keep separate tiles? A credible common dataset for both token price and full cost-to-score, with comparable score, models, date ranges, and uncertainty, could support a linked multi-meter chart—but not simple value splicing.

How much of Epoch’s measured decline transfers to longitudinal, quality-checked real work rather than benchmark questions? The report does not establish a general conversion.
What is the cost of switching models, integrating them, retrying failures, and reviewing answers for a particular buyer? The report explicitly focuses on a cheapest frontier, not a complete enterprise bill.
Does the same slope persist beyond the September 2026 observation window? Not established; changes in test difficulty, model supply, energy, or deployment costs could alter it.
How much of the decline is hardware efficiency versus algorithms, pricing strategy, competition, and more efficient reasoning? The report does not causally decompose the aggregate rate.