AI is getting 13 times cheaper every year

The price of getting AI to solve a task at a certain level is falling extremely fast. That is what a report from the research institute Epoch AI shows, published in September 2026 by Luke Emberson and David Roodman.

On average, the price of a given performance level has fallen about 47 percent every quarter for the past three years. That corresponds to it becoming roughly 13 times cheaper per year.

An example: 725 times cheaper

Epoch cites OpenAI's model o3, which came out on January 31, 2025. It could score 75 percent on GPQA Diamond, a multiple-choice test with PhD-level questions in physics, chemistry and biology. It cost an average of 30 cents per question.

Just under 18 months later, OpenAI released the model GPT-5.6 Luna, which scored just as high for about 0.04 cents per question. That is a price drop of 725 times. Epoch compares it to a new car falling in price from $50,000 to $69.

The pace varies. On game-based tasks, the price falls 39-43 percent per quarter, and on math tasks 50-52 percent. The decline typically slows over time: right after a new level is reached, the price falls about 66 percent per quarter, but two years later only about 32 percent.

One possible explanation is that AI companies can briefly charge a premium for a new level before competition and better technology push the price down.

Caveats

Epoch stresses that the figures should not be read as entirely precise. Companies may have trained models specifically for these tests (called "benchmaxxing"), so the progress looks bigger than in real-world work. A good score on a test is also not the same as useful work.

The analysis also looks at the cheapest model that can reach a given level. That assumes a user who constantly switches to the most economical model. Ordinary users switch less often and therefore don't get quite the same savings. The data also covers only three years and is not complete.

What does it mean for you?

For companies and users, it means tasks that are expensive to solve with AI today can quickly become cheap. However, the price drop applies to a fixed level of performance. The newest and best models have not necessarily become cheaper.

At the same time, there is an opposite trend: the AI boom is pushing up prices for chips, power and labor such as electricians. The cheap output thus requires expensive inputs.

Epoch has also previously measured large price drops, including 40 times per year to reach GPT-4 level on science questions. The new report suggests the trend has lasted a long time and is broad.