AI inference costs drop 47% every quarter... Plunged to 1/725 in 18 months.

The amount of money spent on solving the same problem decreases by nearly half every quarter.

Doctor-level problem solving cost: 30 cents → 0.04 cents... 1 in 725 in 18 months

(Seoul = Yonhap News) Reporter Oh Ji-eun = Analysis has shown that the 'inference cost' required for artificial intelligence (AI) to solve the same level of problems and come up with an answer is falling by one-thirteenth every year.

Due to AI competition, the cost of infrastructure such as graphics processing units (GPUs), power, and data centers is soaring, but the 'AI price paradox' is emerging, where the price for users to use AI with the same performance is falling so quickly that it is difficult to find in past major technologies such as batteries and computer computing power.

◇ AI inference cost decreases to 1/13th every year... Decline 4 times faster than DNA analysis

Epoch AI, a non-profit research institute, announced on the 4th that in its recently released 'Plummeting Cost of Inference' report, the cost of implementing the same level of AI performance after 2023 has fallen by about 47% every quarter.

Converted to one year, it falls by one-thirteenth (a 13-fold decrease) every year.

The report pointed out that this price decline is at a record pace rarely seen in human industrial history.

When compared to past technologies considering the inflation rate, the rate of decline in AI inference costs is 4 times faster than human genome (DNA) analysis technology, 6 times faster than the development of computer computing power, and 18 times faster than lithium-ion batteries.

Compared to the rate at which electricity prices fell over the 100 years before 1973, this is a whopping 54 times steeper.

This sharp drop in cost is also confirmed in actual performance benchmarks.

The inference model 'o3' introduced by OpenAI in January 2025 cost an average of 30 cents (about 400 won) per question to record a 75% correct answer rate in GPQA Diamond, a doctoral level exam in physics, chemistry, and biology.

However, 'GPT-5.6 Luna', released just a year and a half later, cost only 0.04 cents ($0.0004) to achieve the same score on a test of the same difficulty.

In just 18 months, the price of renting the same level of brain power has fallen to 1/725.

The report said, “This is a drop similar to the price of a new car priced at 50 million won falling to 70,000 won in a year and a half,” and added, “No general-purpose technology in history has ever become cheaper in such a short period of time.”

By field, the cost of solving math problems showed the steepest decline, falling 5,052% per quarter (1,619 times per year), while chess and game-based puzzles fell 3,943% per quarter (710 times per year).

◇ When a new model comes out, prices fall faster... Big Tech’s ‘monopoly profits’ are also shortening

What's notable is that the price collapse is most severe shortly after the new top-performance model first hits the market.

Immediately after a new model surpassed the best performance, competing models quickly imitated and caught up, and the cost of realizing that level of performance plummeted by 66% per quarter (75 times per year). After two years, the decline appeared to have calmed down somewhat to 32% per quarter (4.7 times per year).

The report explained that this means that even if big tech companies pour enormous capital and power to develop cutting-edge models, the period in which they can enjoy excess profits by charging high fees based on their technological superiority is only a moment.

In fact, a fierce chase is taking place in the market between the closed proprietary model and the open source camp, rapidly driving down the unit price relative to performance.

The report analyzed, “Due to the AI ​​boom, input costs such as graphics processing units (GPUs), power, and data center infrastructure are skyrocketing, but a paradox is occurring in which the market price of inference services, which are the actual output, is plummeting.”

He added, "While companies and consumers can now easily introduce high-performance AI at a much lower cost than before, it is becoming increasingly difficult to maintain short-term monopoly profits as an extreme price reduction race continues among model developers."