The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Yes—but the clearest evidence is about the cost of reaching a fixed level of AI benchmark performance, not every AI product or expense. Epoch AI’s September 2026 analysis estimates that this benchmark-adjusted inference cost fell about 47% per quarter from 2023 across five benchmarks. That is roughly a 13-fold reduction per year on the study’s fitted trend. The result is unusually fast compared with several historical technology price series the authors examined, but it does not mean every AI service is getting cheaper at that rate.
What does “AI is getting cheaper” mean?
The strongest evidence measures inference: the expense of querying a trained model to achieve a specified level of performance. Epoch AI estimates a cost-performance frontier for each benchmark and target score—the least expensive available model-and-run combination that meets or exceeds that score.
This fixed-performance approach matters. Comparing only the price per token of two unlike models can be misleading: a newer model may charge more per token yet need fewer tokens, or a less costly run, to reach a particular score. Stanford HAI’s 2025 AI Index also describes fixed-performance comparisons as more informative than directly comparing prices for newer and older models. Its price series combines Artificial Analysis and Epoch AI API-pricing data, weights input tokens three to one against output tokens, and reports U.S. dollars per million tokens.
Free tools Windows power users keep installed
One-click scans. No signup required.
How much has AI inference cost fallen?
Epoch AI’s September 2026 analysis estimates a decline of about 47% per quarter since 2023 in the cost of achieving a given performance level across five benchmarks spanning mathematics, hard sciences and games of skill. The authors express that as about 13 times lower per year. This is a fitted trend across benchmark frontiers, not a guaranteed discount for a particular user or API.
#1 Best Overall
A separate historical example comes from Stanford HAI’s 2025 AI Index: the reported inference price for models scoring at a GPT-3.5-equivalent level on MMLU fell from $20 to $0.07 per million tokens between November 2022 and October 2024—more than a 280-fold decrease in roughly 1.5 years. This is a past fixed-performance comparison, not a current price quote. Stanford also reported a fall from $15 to $0.12 per million tokens for models scoring above 50% on GPQA between May and December 2024.
Does the decline apply equally to every task?
No. Epoch AI estimates quarterly declines of 39–43% for game-based puzzles and 50–52% for math problems. In its average across the five primary benchmarks, the cost of newly achieved state-of-the-art performance fell 66% per quarter; two years after a performance level first reached the frontier, the estimated decline was 32% per quarter. In other words, the rate varies both by task and by how recently a capability became state of the art.
Stanford HAI, reporting earlier Epoch AI estimates, gave a task-dependent range of nine to 900 times per year. That wide range is a reminder that any headline rate depends on the task, target performance, date and estimation method.
How does the study estimate cost at a fixed score?
Models can be run with different reasoning settings or token budgets, changing both their cost and benchmark score. Rather than repeatedly running every model at every possible budget, Epoch AI uses a procedure developed by the federal Center for AI Standards and Innovation (CAISI): high-budget benchmark-run transcripts help estimate how performance would change under tighter budgets. For open-weight models without a dedicated inference API, the analysis estimates cost using rented hardware; Epoch reports that, for five models it validated, these estimates differed from API pricing by less than 30%.
The resulting frontier is an estimate of the cheapest way to reach each benchmark target among the models and runs represented in the data. It assumes a user can switch actively to the most cost-effective option for a given target, which many people do not do for every task. The five primary benchmark series begin in 2023; Epoch’s suggestion that a similar pace may extend back to commercial large-language-model inference beginning in November 2021 rests on coarser earlier evidence, so the measured result is best described as applying since 2023.
Is AI getting cheaper faster than other technologies?
In Epoch AI’s comparison, the estimated AI benchmark-performance cost decline is faster than the declines in four selected historical price series. The periods and rates differ by technology:
| Technology and measure | Period examined | Annual decline rate stated by Epoch AI |
|---|---|---|
| AI cost to reach benchmark performance | Since 2023 in the five primary benchmark series | About 13× lower per year, equivalent to the estimated 47% quarterly decline |
| DNA sequencing | 2001–2025 | 1.84× per year |
| Compute | 1940–2001 | 1.51× per year |
| Lithium-ion batteries | 1991–2024 | 1.16× per year |
| U.S. residential electricity | 1892–1973 | 1.05× per year |
Epoch describes the AI decline, in log-point terms, as four times faster than DNA sequencing, six times faster than compute, 18 times faster than batteries and 54 times faster than residential electricity. These comparisons are not like-for-like: benchmark performance, sequencing, computing, battery prices and electricity are different outputs measured in different ways and over different periods. Epoch itself calls the exercise apples to oranges. It supports a qualified claim about the specific series compared, not proof that AI is becoming cheaper faster than every technology in history.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What is not established by these falling costs?
- Every AI price is falling at the same rate. The study measures the least cost to reach benchmark targets, not a universal price list for subscriptions, APIs or workloads.
- Frontier services are always inexpensive. Stanford HAI notes that state-of-the-art models can remain more expensive than smaller alternatives. The cost of reaching an established capability target can fall even while access to the newest frontier model costs more.
- Training and infrastructure costs are falling at this rate. The evidence concerns inference. It does not establish a matching decline in model training, chips, power, labor or data-center costs.
- Benchmark results translate directly into cheaper useful work. Benchmarks are proxies, and developers may optimize for tests. A benchmark score does not guarantee equivalent performance, value or savings in an ordinary task.
What should an AI user take from the numbers?
The practical takeaway is that a given level of measured capability may be attainable at much lower inference cost than it was before, and the available evidence shows a particularly rapid decline across Epoch AI’s five benchmarks since 2023. But the savings for a real workload depend on its quality target, the model and settings chosen, token use, API pricing and whether the user changes models as costs and capabilities shift. The headline rate describes a benchmark-based frontier—not a promise about an individual bill.
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.
