Epoch AI reports 13-fold annual decline in AI costs as spending nears $1T

AI is getting cheaper at a pace that makes most technology history look sluggish. According to a new Epoch AI report, the cost of hitting a fixed level of AI performance has fallen roughly 13-fold per year since 2023.

Meanwhile, the companies building AI are not saving that money. US and Chinese firms are projected to pour nearly $1 trillion into AI infrastructure this year, which is a strange way to respond to a product that keeps getting cheaper.

The plunging price of thought

The report, published September 22, 2026, carries the title “The Plunging Price of Thought.” It was written by Luke Emberson and David Roodman.

The core finding: across the benchmarks Epoch monitors, the expense of reaching a given performance level dropped by approximately 47% per quarter. Compounded over a year, that works out to approximately 13 times cheaper.

Epoch’s sharpest example comes from the GPQA Diamond benchmark. In January 2025, OpenAI’s o3 model scored 75% on it at a cost of roughly $0.30 per question.

A later model, GPT-5.6 Luna, matched that same 75% score for $0.0004 per question. That is a 725-fold drop in cost for an identical result.

The report also finds the decline is not uniform over time. Costs fall hardest right after a capability first becomes state-of-the-art, averaging around 66% per quarter.

That pace slows as a capability matures. Two years after reaching the frontier, the decline tapers to approximately 32% per quarter.

Faster than DNA, chips, and batteries

To put the numbers in perspective, Epoch compared AI’s cost curve against other transformative technologies. DNA sequencing fell four times slower than AI performance costs. General computing declined six times slower. Lithium-ion batteries dropped eighteen times slower.

One important caveat sits inside the headline figure. The 13-fold annual decline bundles together several forces: better hardware, competition among providers, and algorithmic improvements. Strip out everything except the algorithms themselves, and Epoch estimates pure algorithmic efficiency improves at around 3 times per year.

Cheaper AI, bigger bills

Falling unit costs have not slowed spending. Bloomberg forecasts that US and Chinese hyperscalers and neoclouds are set to invest nearly $1 trillion in AI-related capital expenditures in 2026, nearly double what they spent the previous year.

Both groups are betting that cheaper AI will not mean less demand for AI infrastructure. If anything, they appear to be wagering that lower prices will unlock far more usage.

What this means

For the companies spending that nearly $1 trillion, the Epoch findings cut both ways. Cheaper inference makes AI viable for far more applications, which supports the case for building more data centers. But it also raises questions the report itself flags: will workload demand actually grow fast enough to fill all that new capacity, and what will utilization rates look like if each task requires dramatically less compute than before?

Because pure algorithmic progress runs closer to 3 times per year, a large chunk of the 13-fold figure depends on hardware gains and competitive pricing.

For AI model providers, any lead in a given capability may be short-lived, since the cost of matching it collapses fastest right after it debuts, at roughly 66% per quarter, before decelerating to around 32% two years later.