Back to Devoured - September 23, 2026
DEVOURED
The plunging price of thought

The plunging price of thought

Tech Epoch AI
The cost of AI intelligence has collapsed by approximately 47% per quarter, suggesting it is becoming the fastest-devaluing technology in history.
What: Researchers at Epoch AI found the cost of a given level of AI performance has fallen by 13x annually since late 2021. Price drops are most aggressive for state-of-the-art models, which decline 66% per quarter shortly after release before stabilizing to slower decreases.
Why it matters: Rapid cost deflation implies that high-performance AI is shifting from a scarce resource to a commodity, forcing AI labs to pivot toward building defensible application layers rather than relying on model-training premiums.
Deep dive
  • The cost of AI performance has fallen at a rate of 47% per quarter since November 2021.
  • Price declines for state-of-the-art models are significantly faster than for legacy technology like electricity or lithium batteries.
  • Performance benchmarks in math have seen faster cost declines (50–52% per quarter) than game-based puzzles (39–43% per quarter).
  • The study utilized the CAISI methodology to model cost-performance frontiers across various reasoning budgets.
  • Researchers warn that 'benchmaxxing' (optimizing models for specific test sets) may artificially inflate capability perception.
  • Future cost models likely face diminishing returns as hardware and energy costs form a price floor.
  • The analysis highlights that closed-model labs currently lead in near-frontier price competition.
Decoder
  • Pareto frontier: The set of configurations where one variable cannot be improved without worsening another; in this context, the lowest possible cost for a given performance level.
  • Reasoning models: LLMs (like OpenAI's o-series) designed to spend more compute tokens on internal 'thought' chains before generating final answers.
  • Stochastic Frontier Analysis (SFA): An econometric method used to estimate production efficiency by modeling the distance between observed output and the best-possible theoretical output.
Original article

Full article content is not available for inline reading.

Read the original article →

Explore all 56 articles in today's digest

Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi