The price of AI thinking falls 13x a year

Reaching the same AI benchmark score gets about 13x cheaper every year.

The numbers

1. Five benchmarks, since 202313x

   cheaper per year to reach the same score.

   About 47% cheaper every quarter.

   Source: Epoch AI, The plunging price of thought, 22 Sep 2026

2. One score: 75% on GPQA Diamond, on Epoch’s guess-adjusted scale

   725x cheaper in under 18 months

   Jan 2025, o3$0.30per question, Epoch’s estimate

   2026, GPT-5.6 Luna$0.0004per question

   Source: Epoch AI, 22 Sep 2026

3. How much faster AI’s price fell than

   Faster than any price drop we’ve seen

   • Electricity, 1892 to 197354x
   • Lithium batteries18x
   • Computing6x
   • DNA sequencing4x

   Compared by rate of decline.

   Source: Epoch AI, 22 Sep 2026

4. Price drop per quarter, same capability

   The drop is fastest at the start

   When it’s brand new-66%per quarter, about 75x a year

   Two years later-32%per quarter, about 4.7x a year

   Source: Epoch AI, 22 Sep 2026

5. The catch3-18x

   more expensive per year to run frontier models.

   A separate study finds a fixed level gets 5 to 10x cheaper a year, while the newest models cost more to run.

   Source: arXiv 2511.23455, Mar 2026

Plan for prices that keep falling

1. Re-check your model every quarter. Last quarter’s frontier is this quarter’s cheap option.
2. Use the cheapest model that passes your evals. Keep the frontier model for the hard 10%.
3. Don’t build a business on today’s price. Margins built on token cost move every few months.

Sources

1. Epoch AI, The plunging price of thought, 22 Sep 2026
2. Gundlach, Lynch, Mertens, Thompson, arXiv 2511.23455 (v2), 23 Mar 2026

When did you last re-price your AI stack? Tell me on LinkedIn or Instagram.