Vol. XXX · The Research Frontier Plate 7 of 11efficiency, algorithmic progress, price per token, cost of capability, Epoch AI, compoundingCE

Doing More With Less.

Yesterday’s frontier becomes cheap quickly; today’s frontier keeps getting dearer.

This volume describes the field as of October 2026. Claims are marked established (replicated or long-standing), reported (from a single lab or report, not independently reproduced) or contested (actively disputed). Two measured trends compound. Algorithmic progress: Ho and colleagues (NeurIPS 2024) estimated that the compute needed to reach a fixed pretraining performance halved about every 8 months over 2012–2023 (90% interval 2–22 months; established as an estimate with wide uncertainty). Price of inference: Epoch AI found the price of reaching a fixed benchmark score fell between 9× and 900× a year depending on the threshold, median about 50× (reported, with the fastest declines the most recent and least likely to persist).

Hardware adds its own improvement (plate XIII.1), and the volumes on efficient inference (XXIII–XXIV) explain the mechanisms. The workshop compounds the chosen yearly rates over the chosen number of years to show the cost of a fixed capability falling, against the cost of each year’s frontier run rising.

Below: relative cost of a fixed capability and of the frontier training run over time, on a logarithmic scale.

What you will learn
Workshop
Fixed capability: inference price falling by the chosen factor per year. Pretraining compute for a fixed loss: halving every 8 months (2.8× a year). Frontier run cost: growing by the chosen factor per year (spending, not efficiency). All relative to year 0. Illustrative compounding of published rates, not a forecast.

Cheaper capability, dearer frontier

Halving compute every 8 months is about 2.8× a year: over three years, a fixed pretraining loss needs about 23 times less compute. Inference prices for a fixed score fell faster still — at the 50× median, three years would be a 125 thousand-fold fall, at the 9× low end 729-fold; these rates are measured over a short, unusual period and the authors themselves expect them to slow. Either way, last year’s frontier capability becomes cheap quickly, which is why small open models keep matching benchmarks that were frontier results a year or two earlier.

The frontier run itself gets more expensive at the same time, because developers spend the efficiency gains on more capability rather than on cheaper copies of the old one; Cottier and colleagues estimated frontier training costs growing about 2–3× a year (reported). Both statements are true together: what a model can do for a given price rises steeply, and the price of the best model rises too. Forecasts that extrapolate either trend alone mislead.

Efficiency gains are uneven across abilities: knowledge-heavy tasks have resisted compression into small models more than benchmark-style reasoning, so “the same capability” is itself a simplification.

Terms

Algorithmic progress
Less compute needed for the same result through better methods.
Price of capability
The cost of reaching a fixed level of performance.
Frontier run
The largest training run of its time.
Compounding
Multiplying a yearly rate over several years.
Capability gets cheaper fast at a fixed level while frontier runs get dearer; both trends hold at once.

Sources

  1. Ho et al., “Algorithmic Progress in Language Models”, NeurIPS 2024.
  2. Epoch AI, “LLM inference price trends”, 2025.
  3. Cottier et al., “The rising costs of training frontier AI models”, arXiv:2405.21015, 2024.