Recommended Free Tools
By Epoch AI’s estimate, the cost of reaching a given score on five AI benchmarks fell about 47% per quarter—roughly 13-fold per year—between 2023 and 2026. That is faster than the historical price declines Epoch compares it with, including compute. But this is a benchmark-performance cost frontier, not a measure of every AI subscription, API call or real-world task. The result is striking; the phrase “the price of AI” needs that qualification.
What does “the price of AI” mean here?
Epoch AI’s September 22, 2026 report, The plunging price of thought, tracks the cheapest model it could identify that reached specified scores on five benchmarks spanning mathematics, hard sciences and games of skill. It asks how the cost of reaching a fixed level of benchmark performance changed over roughly three years. This is a cost frontier: it assumes a buyer can search among available models and choose the least expensive one that meets a target.
As an Amazon Associate I earn from qualifying purchases.
It is not an average price index for AI products. It does not say that every service, subscription or workflow became 13 times cheaper, nor does it measure total spending or the economic value of AI. A benchmark score is also only a proxy for useful work, and a system can be especially good at a benchmark without being the best choice for a particular user.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsHow fast did benchmark-performance costs fall?
Epoch estimates an average decline of about 47% per quarter, equivalent to roughly 13-fold per year, across its five benchmarks since 2023. The report summarizes its finding this way: “The cost of achieving a given level of AI performance has fallen about 47% per quarter, faster than for any other transformative technology in history.” That estimate is Epoch AI’s result, not a universal measure of AI prices.
#1 Best Overall
The decline differs by benchmark type and by how recently a performance level became possible:
| Measure | Epoch AI estimate | What it describes |
|---|---|---|
| Across five benchmarks | About 47% per quarter; about 13× per year | Average decline in the cheapest identified cost of reaching specified benchmark scores since 2023 |
| Game-based puzzles | About 39–43% per quarter | Cost declines on the report’s game-based benchmarks |
| Math problems | About 50–52% per quarter | Cost declines on the report’s math benchmarks |
| Newly achieved state-of-the-art performance | About 66% per quarter; about 75× per year | Average across five benchmarks when the target is at the frontier |
| Same performance, two years later | About 32% per quarter; about 4.7× per year | Average across five benchmarks after a target has been achievable for two years |
The frontier comparison helps explain why “AI costs are in freefall” can sound more sweeping than it is. The fastest decline applies to the newest performance frontier; a capability that is no longer cutting-edge can still get cheaper quickly, but Epoch’s average rate is lower two years after that level first appears.
What does the comparison with Moore’s Law show?
Epoch compares rates of price decline, not identical products or like-for-like units. Its compute comparator is a historical price series from 1940 to 2001, rather than a claim that Moore’s Law itself governs current AI prices. Moore’s Law is shorthand for a historical trend in computing; Epoch’s direct comparison is with the price of compute over a particular period.
Free tools Windows power users keep installed
One-click scans. No signup required.
Rank #2
| Technology and period in Epoch AI’s comparison | Historical annual price-decline multiplier |
|---|---|
| US residential electricity, 1892–1973 | 1.05× |
| Lithium-ion batteries, 1991–2024 | 1.16× |
| Compute, 1940–2001 | 1.51× |
| DNA sequencing, 2001–2025 | 1.84× |
| AI benchmark-performance costs, Epoch AI estimate since 2023 | About 13× per year |
On those series, Epoch characterizes its AI decline rate as about four times faster than DNA sequencing, six times faster than compute, 18 times faster than lithium-ion batteries and 54 times faster than electricity. These are comparisons of estimated rates across different periods, units and kinds of products—not proof that AI is the same kind of good as electricity, batteries or sequencing.
Epoch explicitly describes the comparison as “apples and oranges.” AI’s measured basket shifts among models, services, benchmarks and performance levels. The other historical series cover more homogeneous products. The time series for AI is also short, incomplete and noisy, and does not include every model-and-benchmark combination.
What does a concrete benchmark example look like?
Epoch estimates that OpenAI o3, released January 31, 2025, could score 75% on GPQA Diamond at an average cost of $0.30 per question. The report says GPT-5.6 Luna later reached the same score for $0.0004 per question, a 725-fold reduction in under 18 months. These are Epoch’s estimates for that benchmark comparison; they are not general per-question prices for using either model, and they do not establish which model is preferable for other tasks.
Rank #3
Does cheaper AI mean users or companies will spend less?
Not necessarily. Epoch’s frontier describes what someone might pay by finding the cheapest model that clears a target. Users may stick with an existing tool, lack access to a lower-cost option, or need quality, reliability, speed or features that the benchmark comparison does not capture. Nor does a lower provider cost guarantee an equivalent reduction in customer prices.
Other studies illuminate different parts of the market, but their figures should not be merged with Epoch’s benchmark rate:
- Market prices and model availability: A 2026 Journal of Economic Perspectives article by Mert Demirer, Andrey Fradkin and Nadav Tadelis describes a rapidly expanding inference market, frequent turnover among leading models and creators, and no single model dominating every use case. Its analysis reports a 1,000-fold decline in the price of intelligence and prices about 90% lower for open-source models than comparable closed-source models. Those are results from a separate market analysis, with a different framing from Epoch’s cost-to-benchmark frontier.
- Provider costs in a future scenario: Gartner’s March 25, 2026 forecast projects provider inference costs for a one-trillion-parameter LLM to be more than 90% lower in 2030 than in 2025. This is a forecast, not an observed decline; Gartner says the scenario varies with semiconductor assumptions. It also cautions that lower unit costs may not fully pass through to enterprise customers.
Gartner’s warning matters for AI agents and other token-intensive uses: agentic models can require 5–30 times more tokens per task than a standard chatbot, according to the forecast’s accompanying discussion. If usage or tokens per task rise faster than unit costs fall, a company’s total inference bill can grow even while each unit of inference becomes cheaper. Gartner analyst Will Sommer put the distinction this way in the March 25, 2026 release: “Chief Product Officers (CPOs) should not confuse the deflation of commodity tokens with the democratization of frontier reasoning.”
How should you compare AI options for a real task?
The cheapest benchmark-capable model is not automatically the lowest-cost way to finish your work. Compare options on the task you actually need done, rather than assuming the headline decline applies to your bill.
- Test output quality against your task’s requirements, including accuracy and any need for citations or structured output.
- Compare cost per completed task, not just a token rate; include prompts, retries and any extra steps in a workflow.
- Check latency, availability and access in your region, as well as whether the quoted figure is a provider cost or the price charged to you.
- For agents, account for the number of model calls and tokens used across the whole task; a lower unit price can be offset by greater usage.
The cited studies do not identify one best provider or model for every use case. The practical question is whether a specific option meets your quality and access needs at a lower total cost for your own task.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
