Primary source AIArticle·23 September 2026·6 min read

The Price of Artificial Thought Is Falling 13x a Year, a Historic Record

Epoch AI measured how fast a given level of AI performance becomes cheap. No disruptive technology has ever gotten cheaper this quickly.

ShareWhatsApp X LinkedIn Facebook

Save to my selection

The Price of Artificial Thought Is Falling 13x a Year, a Historic Record

Source : Luke Emberson (Epoch AI) · Epoch AIView original ↗

In brief

According to Epoch AI, the cost to reach a given performance level across five benchmarks (math, science, games) has dropped roughly 47% per quarter since 2023, a factor of 13x per year. That's far faster than electricity, batteries, compute, or DNA sequencing. The drop is even more brutal right after a performance level becomes state of the art, which says a lot about how fragile lab margins really are.

🍺 Bar-stool version

Answering a PhD-level physics multiple-choice question cost 30 cents in early 2025; eighteen months later, it costs four hundredths of a cent. It's like your brand-new car going from $50,000 to $69, and the dealer thinking that's totally normal. Electricity took eighty years to do much less, and it never learned to play chess. Except cheaper doesn't mean less spending overall: once something becomes nearly free, people end up ordering a million of it.

Key takeaways

1
Since 2023, the cost of a fixed AI performance level has fallen by roughly 47% per quarter, or 13x per year, averaged across five benchmarks: FrontierMath (tiers 1-3), OTIS Mock AIME, GPQA Diamond, Chess Puzzles, and Mystery Game Puzzles.
2
Math drops fastest (50–52% per quarter, 16–19x per year), game puzzles slowest (39–43%, 7–10x per year).
3
Flagship example: 75% on GPQA Diamond cost about $0.30 per question with o3 (January 2025), then $0.0004 with GPT-5.6 Luna less than 18 months later — a 725x drop.
4
Right after a performance level becomes state of the art, cost falls 66% per quarter (75x per year); two years later, the decline slows to 32% per quarter (4.7x per year).
5
In log points, LLM inference has fallen 4 times faster than DNA sequencing, 6 times faster than compute, 18 times faster than lithium-ion batteries, and 54 times faster than electricity.
6
The method reconstructs each model's full cost-performance curve from benchmark transcripts, following a procedure from the Center for AI Standards and Innovation (CAISI).
7
The authors themselves list the limitations: benchmaxxing, the gap between benchmarks and useful work, real users who don't constantly switch models, and only three years of data.

An output price collapsing while inputs skyrocket

The AI boom is driving up the price of what it consumes: chips, electricity, and even electricians' labor.

Measuring real cost, not token price

Previous work reasoned in terms of price per token: Guido Appenzeller (a16z) found 10x per year, an earlier Epoch analysis from March 2025 found between 9 and 900x depending on the benchmark.

Statistical models presented as approximate

The object of study is the Pareto frontier: the cheapest model able to reach each performance level at a given date.

The state-of-the-art premium doesn't last

On three of the five main benchmarks (AIME, FrontierMath, GPQA Diamond), cost falls fastest right after a performance level is first achieved.

A decline with no historical equivalent

Epoch considers it plausible that this pace has held since November 2021, when GPT-3's API opened commercially.

The limits, stated by the authors themselves

First risk: benchmaxxing, training specifically targeted at known benchmarks.

Why it matters

This report puts a solid order of magnitude on a diffuse intuition: intelligence at a given level is becoming a commodity at an unprecedented pace.

Older than 30 days: the archive

This article is more than 30 days old

A free account opens articles from the last 30 days in full. The full archive is in the Pro plan.

Create my account

Already a member? Log in

Compare plans →

#ai #llm #costs #inference #benchmarks #epoch ai

Original source

The plunging price of thought

Luke Emberson (Epoch AI)

Open the article ↗

For you

Put it to work on your sources.

Free: a month of articles and three sources of your own. Pro: the whole archive and your sources, from €8/month.

Create my account →

For your team

The same machine, on your topics.

A space in your colours, your watch angles, your curators. Pilot open to three companies.

Request pilot access →