The research behind Two Prices for the Same Intelligence. This dossier distinguishes a named token-price observation from a modeled performance-cost series, preserves the GPQA Diamond cost-record observations, and sets out methodology, limits, and unanswered questions. It does not reproduce Epoch AI’s paper; read the original report for the full analysis.
- 01 — What Altman actually claimed
- 02 — What Epoch measured
- 03 — The GPQA cost-record series
- 04 — Reconciliation and boundaries
- References and sources
- Question ledger
The broad signal agrees: a fixed level of AI performance has become much cheaper. The units do not: Altman’s example is a price per token for a pair of models; Epoch estimates the total cost to attain a test score across models and reasoning budgets. The named 10× and estimated 13× rates should therefore be shown alongside each other, not spliced into one data series.