Simple Explanation (ELI5)

Think of AI like a cake: the price tag says it’s cheaper, but if you check how big the cake really is, you see it’s much smaller. So the price looks lower, but the real value has dropped a lot.

Key Advantages & Current Limits

🟢 Pros: Open, free data enables anyone to verify and reproduce the study at zero cost 🟢 Pros: The quality‑adjusted index reveals a much larger price drop (0.73% vs 0.10%), giving a clearer picture of market dynamics 🔴 Cons: The matched‑model approach relies on how prices are linked to benchmarks, which can introduce bias if the benchmarks are not fully representative 🔴 Cons: Reasoning models consume more tokens while token prices fall slowly, causing a gap between price per token and price per completed task

Future Outlook & Takeaways

In the next 3‑5 years, quality‑adjusted indices will make the AI market more transparent, letting businesses assess true service value. This will boost competition, reduce concentration among large providers, and raise productivity through more realistic pricing.

How Much Does AI Really Cost?

Article

100%

Artificial intelligence: a new way to measure how much its services cost.
The published prices for AI services have fallen since 2024.
The speed of this decline depends on how we measure it.
A team of researchers collected 21,024 price observations from 86 providers and 3,208 different models.
They then linked them to 4,605 benchmark scores, creating a hidden quality index.
The “matched-model” method, used by statistical agencies for software, showed a price drop of 0.10% per year.
But the quality‑adjusted index fell by 0.73%, hiding 87% of the real decline.
Reasoning models consume more tokens, while token prices fall slowly.
Thus, the price for buyers and sellers diverge.
A pre‑registered validation study checks the data and confirms the results.
Everything is freely reproducible at zero cost.
This study shows how crucial it is to consider quality when measuring the AI market.
Understanding these numbers helps us see competition, concentration, and productivity more clearly.
Young people interested in technology and economics can follow these studies to discover how AI is changing the world of work.

Bonus

100%

This paper introduces a quality‑adjusted price index for the AI inference market, addressing the shortcomings of conventional measurement approaches. Researchers compiled 21,024 publicly listed price observations from 86 providers and 3,208 distinct models, linking them to 4,605 benchmark scores sourced from repositories such as HELM, BIG‑bench, and ML‑Perf. By applying a latent factor model to benchmark response patterns, they constructed a hidden quality dimension that replaces traditional hedonic product attributes. Using the matched‑model technique—commonly employed by statistical agencies for software—they found a raw price decline of 0.10% per year. However, the quality‑adjusted index revealed a steeper drop of 0.73%, indicating that current methods miss 87% of the actual price decrease. A pre‑registered validation study (available on OpenScience) flagged contaminated benchmarks; removing these shifts the index by 0.49 log points annually while preserving model rankings at 99.8% stability. Reasoning models (e.g., GPT‑4, Claude) have increased token consumption faster than token prices have fallen, leading to a divergence between per‑token and per‑task pricing for buyers and sellers. These findings directly impact how competition, market concentration, and productivity are measured, as existing statistics underestimate price dynamics. All data, code, and results are openly shared on GitHub and Zenodo, ensuring full, cost‑free reproducibility.