← AI Switchboard
AI Switchboardby Waggle
COST · September 24, 2026
Sep 22

Epoch: the cost of a fixed level of AI performance falls about 47% a quarter

Epoch AI measures the cheapest model able to hit a given benchmark score, not the price of the best model available. On that measure cost falls roughly thirteen-fold a year.

Epoch AI published “The plunging price of thought” on Monday, and the quantity it measures needs stating precisely, because the sloppy version of this finding is wrong. Epoch is not saying AI is getting cheaper. It is saying that the cost of reaching a given level of performance falls about 47% per quarter, or roughly 13x per year, since 2023. In Epoch's words, “we focus on the frontier — the absolute cheapest model capable of any given level of performance.” That is the cheapness frontier, not the capability frontier.

The rate varies by domain. Maths falls fastest, 50-52% per quarter, with FrontierMath tiers 1-3 specifically at 53.1% by Epoch's model-free estimate. Game-based puzzles fall slowest, 39-43%. The worked dollar example is GPQA Diamond at the 75% threshold: OpenAI's o3, released 31 January 2025, cost about 30 cents a question; GPT-5.6 Luna, just under eighteen months later, cost $0.0004 — a 725-fold drop.

Epoch attaches four caveats of its own, and they matter as much as the headline. Benchmaxxing. Benchmarks as an imperfect proxy for useful work. Barely three years of data. And the one that decides whether any of this reaches your invoice: the full saving accrues only to a user who relentlessly switches to the cheapest adequate model, which “real users do not” do. Epoch also says plainly that its bottom-line numbers “should not be read as exact.”

One thing this report does not say, despite being cited for it: nothing here shows the blended price of the frontier falling. Epoch does not address that question at all.

  • Confirmed Cost to reach a given level of performance has fallen about 47% per quarter, or 13x per year, since 2023 — the falling quantity is cost at fixed capability, not the price of the frontier. Epoch AI
  • Confirmed Maths falls fastest at 50-52% per quarter (FrontierMath T1-3 at 53.1%); game-based puzzles slowest at 39-43%. Epoch AI
  • Confirmed Worked example, GPQA Diamond at the 75% threshold: OpenAI o3 at about $0.30 a question versus GPT-5.6 Luna at $0.0004 just under 18 months later — 725-fold. Epoch AI
Sources: Epoch AI

Money & marketsModels & releases

This week in the September 24, 2026 edition · front page