Epoch AI: The Cost of AI Performance Falls 13x Every Year

The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13 times per year, the research group Epoch AI says in a report published on 22 September 2026, “The plunging price of thought.”

Epoch says that drop is four times faster than DNA sequencing, six times faster than compute, 18 times faster than lithium batteries and, in the century up to 1973, 54 times faster than electricity. The figures were circulating again on X on 14 October.

What Epoch Measured

Epoch says it analysed a new dataset of five AI performance benchmarks covering mathematics, hard sciences and games of skill over the last three years. Across that time, it finds the price for a given level of performance has fallen about 47% per quarter.

The speed varies by domain, according to the report: slower on game-based puzzles, at 39–43% per quarter, and faster on math problems, at 50–52% per quarter. Epoch says coarser evidence suggests the price of thought has been falling at least this fast since November 2021, when OpenAI fully released GPT-3.

One Example: 725 Times Cheaper

The report gives one worked example. Epoch estimates that OpenAI’s o3, released on 31 January 2025, could reach a 75% score on GPQA Diamond, a PhD-level science exam, for an average cost of 30 cents per question.

Just under 18 months later, it says, GPT-5.6 Luna scored just as well for four hundredths of a penny per question ($0.0004). Epoch calls that “a 725-fold drop in the price of thought in under 18 months.”

Fastest Right After a Capability Debuts

Epoch says the cost of a given level of performance often falls fastest right after that level is first achieved. Averaging across all five benchmarks, cost falls 66% per quarter (75 times per year) for performance that has just debuted as state of the art. Two years later, prices fall half as fast, at 32% per quarter (4.7 times per year).

One possible explanation, the report says, is that when a performance level is first achieved AI companies can briefly charge a premium for it, before competition and technological improvement drive the price down.

The Caveats Epoch Lists

Epoch says its analysis “comes with major caveats.” AI companies may be training their models for some benchmarks, so benchmark gains could outstrip real-world gains, and doing well on a benchmark is not the same as useful work.

It also says it tracks the absolute cheapest model for each performance level, which implies a user who keeps switching to the most cost-effective model; real users do not switch that often, so they do not get quite the same savings. Its data cover barely three years and are incomplete and noisy, and Epoch says the bottom-line numbers “should not be read as exact.”

Where This Meets Today’s Pricing

The report lands as model makers keep repricing. Google launched Gemini 4 Argon at an introductory price that its launch post says rises after 30 November. On 14 October, OpenAI said it will retire GPT-4o and GPT-4.1 from its API and pointed customers to its Luna model.

The Structural Read

As we read it, the 13-times-a-year figure is a frontier number: Epoch says it tracks the absolute cheapest model for each performance level, and that real users do not switch models so often.

As we read it, the steepest part of the curve sits at the top: Epoch puts the decline at 66% a quarter for performance that has just become state of the art, and at 32% a quarter two years later.

As we read it, the report’s own caveats matter for anyone budgeting on it: Epoch says benchmark scores are not the same as useful work and that its numbers should not be read as exact.

Three Implications

AI BUYERS Epoch says users who do not keep switching to the cheapest capable model do not reap quite the same savings.

MODEL MAKERS Epoch’s possible explanation is that companies briefly charge a premium for a new performance level before competition drives the price down.

BENCHMARK READERS Epoch says companies may be training for some benchmarks, so benchmark gains could outstrip real-world gains.

The Business Engineer Lens

This story maps onto the Business Engineer framework The CFO’s Guide to the Token Economy.

The framework puts it this way: “The cost of intelligence is collapsing while the cost of deploying intelligence is exploding.”

As we read it, Epoch’s report quantifies the first half of that sentence for a fixed level of capability; it does not measure what businesses actually spend on AI.

What Is Not Established

Epoch measures the cost of reaching a fixed score, not what businesses actually spend on AI, and not the price of the newest, most capable model. Its own caveats say benchmark scores are not the same as useful work and that the figures are not exact.

The comparison with DNA sequencing, compute, lithium batteries and electricity is Epoch’s; we have not re-run its data, which it says are published on GitHub.

The Bottom Line

Epoch AI estimates that the cost of a fixed level of AI performance has fallen about 13 times a year since 2023, faster than any other technology it compared, with the steepest drops right after a capability first appears and major caveats it lists itself.