Epoch AI estimates that the cost of reaching a given score on selected AI benchmarks has fallen by about 47% per quarter since 2023, or roughly 13-fold a year. Its figures suggest that performance once costing 30 cents per question on one test could be matched for $0.0004 less than 18 months later.
Epoch AI Watch analysis
What happened
In a report summary published on 28 September, Epoch AI describes comparing the cheapest way to reach specified scores across five benchmarks covering mathematics, hard sciences and games. The project says the cost declines are slower on game-based puzzles, at 39–43% per quarter, and faster on maths problems, at 50–52%.
One example compares OpenAI’s o3, which Epoch AI estimates could score 75% on the PhD-level science benchmark GPQA Diamond for an average of 30 cents per question, with GPT-5.6 Luna at an estimated $0.0004 for the same score. Epoch AI calculates that as a 725-fold drop in under 18 months. These are the project’s estimates, not a market-wide price list.
Why it matters
A model’s capabilities are only part of the access story. If a particular level of performance gets dramatically cheaper to obtain, more developers may be able to use it, and existing products may become cheaper to run. That makes the cost of a benchmark result a useful signal, even if it is not the price every customer will pay.
Epoch AI also says the steepest drops tend to come soon after a performance level is first reached: about 66% per quarter initially, slowing to 32% per quarter two years later. That suggests the pace of price-cutting may change as a capability matures, rather than following one endlessly steep curve.
Our read
The eye-catching figure is not a universal discount on “AI”. It is an estimate of the cheapest route to a set score on selected tests. Epoch AI cautions that benchmarks are imperfect proxies for real-world prices and that its figures should not be read as exact. Still, the comparisons make a valuable point: the cost of equivalent benchmark performance can shift much faster than a model name or headline capability suggests.
For anyone budgeting an AI product, the useful takeaway is to compare the cost of achieving the required result, not just the price tag attached to a model. The benchmark is a guide, not a substitute for testing your own workload.
What to watch
- Whether the longer report’s data and code clarify how the cheapest model and price were selected.
- Whether similar cost declines hold on benchmarks beyond maths, hard sciences and games.
- How closely benchmark-based estimates track prices developers actually pay.
- Whether the slower cost decline after a capability matures continues in future data.
Discussion spark: When AI performance gets dramatically cheaper on benchmarks, should developers treat that as a reason to deploy more widely, or wait for evidence that the savings hold on real-world tasks?
Sources and evidence
- How fast is AI getting cheaper? (28 September 2026, 15:33 UTC)
Independent WittyWires tracker for public updates about Epoch AI. Not affiliated with or endorsed by Epoch AI; this is not an official account.