Yes—by one important measure. Epoch AI estimates that since 2023, the cost of reaching a given score on five AI benchmarks has fallen about 47% per quarter, or roughly 13-fold per year. That is a striking decline, but it describes benchmark-adjusted inference costs—not every AI model price, training expense, or user’s bill.
What does “AI getting cheaper” mean here?
The clearest evidence concerns inference: the cost of querying a trained model and achieving a specified level of benchmark performance. Epoch AI builds a cost-performance frontier for each benchmark and target score, estimating the least expensive model-and-run combination that can meet or exceed that score. Its September 2026 analysis covers five benchmarks spanning mathematics, hard sciences, and games of skill. Epoch AI’s analysis describes the approach and its results.
As an Amazon Associate I earn from qualifying purchases.
This is different from asking how much one million tokens costs. A newer model may charge more per token yet need fewer tokens—or solve the task more reliably—to reach a particular score. Comparing the cheapest route to the same capability can therefore be more informative than comparing headline token rates for unlike models.
It is also a frontier measure: it assumes a user can choose the most cost-effective model and configuration for each target. Someone using one fixed model, setting, or API may not see the same decline.
#1 Best Overall
How quickly have benchmark-adjusted inference costs fallen?
Epoch AI’s September 2026 estimate is an average decline of about 47% per quarter since 2023 across its five primary benchmark series, equivalent to roughly 13 times lower per year. The rate differs by task: the study estimates quarterly declines of 39–43% on game-based puzzles and 50–52% on math problems.
The rate also depends on where a capability sits on the frontier. Across the five benchmarks, Epoch estimates that the cost of newly achieved state-of-the-art performance fell about 66% per quarter; two years after that capability first reached the frontier, the estimated rate was about 32% per quarter. In other words, the cheapest way to reach a new performance frontier can improve especially rapidly, while the cost of an already-established level may continue to fall at a slower pace.
These are fitted trends, not a promise of a particular discount for an individual workload. The five primary series begin in 2023 and the dataset is incomplete and noisy. Epoch suggests a similar decline may extend back to commercial LLM inference in November 2021, but that longer history rests on coarser evidence, not the main five-benchmark measurement.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →What earlier price examples show
A separate fixed-performance example in Stanford HAI’s Artificial Intelligence Index Report 2025 illustrates how much prices can change when capability is held roughly constant. Its series combines API-pricing data from Artificial Analysis and Epoch AI, weights input tokens three to one against output tokens, and reports U.S. dollars per million tokens. These are historical observations, not current quotes.
| Capability comparison | Reported change | Period |
|---|---|---|
| Models at a GPT-3.5-equivalent score on MMLU | $20 to $0.07 per million tokens; more than a 280-fold decrease | November 2022 to October 2024 |
| Models scoring above 50% on GPQA | $15 to $0.12 per million tokens | May 2024 to December 2024 |
Stanford’s examples use fixed capability levels, making them more useful than a simple comparison of old and new model prices. They are still specific to the benchmark, score threshold, pricing data, and dates in the report; they should not be read as what an API costs today. Stanford also reports that earlier inference-cost estimates varied widely by task, from nine to 900 times per year, underscoring why a rate needs its task and method attached.
Is AI getting cheaper faster than other technologies?
In Epoch AI’s comparison, the estimated decline in benchmark-adjusted AI inference costs is faster than the selected historical price trends below. The comparator periods and measured outputs differ, so the numbers are context rather than a perfectly like-for-like ranking.
| Technology and measure | Reported decline rate | Period |
|---|---|---|
| AI cost to reach specified benchmark performance | About 47% per quarter, or about 13× per year | Five benchmark series beginning in 2023; Epoch AI analysis published September 2026 |
| DNA sequencing | 1.84× per year | 2001–2025 |
| Compute | 1.51× per year | 1940–2001 |
| Lithium-ion batteries | 1.16× per year | 1991–2024 |
| U.S. residential electricity | 1.05× per year | 1892–1973 |
Epoch expresses its cross-technology comparison as log-point declines, describing the AI rate as four times faster than DNA sequencing, six times faster than compute, 18 times faster than batteries, and 54 times faster than residential electricity. The authors explicitly characterize this as an apples-to-oranges comparison: AI benchmark performance, sequencing, computing, batteries, and electricity are different outputs measured with different price series. The defensible conclusion is that AI inference costs fell exceptionally quickly relative to these four examples—not that a universal measure proves AI is cheaper than every technology in history.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhy can the measured cost fall so fast?
The estimate tracks the least expensive model and run that can meet a target score, rather than the price of a single unchanged product. As models and serving options improve, a target that once required a costly model or configuration may become achievable through a cheaper route. That can lower the frontier cost even if a particular model’s price does not fall.
Best Value
Epoch estimates how performance changes with reasoning settings or token budgets using high-budget benchmark-run transcripts and a procedure developed by the federal Center for AI Standards and Innovation (CAISI), rather than repeatedly running every model at every budget. For open-weight models without a dedicated inference API, it estimates expenses from rented hardware; it reports that its cost estimates differed from API pricing by less than 30% in validation across five models. These choices make the frontier estimate practical, but they also mean it is an estimate built from benchmark runs and cost assumptions—not a universal invoice price.
What the headline does not establish
- That every real-world task is cheaper. Benchmarks are proxies, and benchmark success does not guarantee equivalent usefulness on ordinary work. Developers may also optimize for tests.
- That every AI price is falling. Stanford HAI notes that state-of-the-art models can remain more expensive than smaller alternatives. Lower cost to achieve a fixed capability can coexist with a higher price for frontier services.
- That training or infrastructure costs are declining at the same rate. The evidence here concerns inference. It does not establish trends in model training, chips, electricity, labor, or data-center costs.
- That users automatically capture the frontier savings. The analysis assumes active model switching to the least expensive way to reach a target. Users who do not switch models or tune configurations may experience a different cost curve.
- That the historical superlative is settled for all time. The benchmark record is short, incomplete, and noisy, and the comparison covers a finite set of technologies and unlike measures.
What this means for someone paying for AI
Use the finding as evidence that the cost of a given level of inference capability can fall rapidly—not as a forecast of your subscription or API bill. Your actual cost depends on the task, the quality threshold, model choice, token use, and whether a less expensive model performs well enough. For a practical comparison, hold the task and acceptable result constant, then compare the total cost of achieving that result rather than price per token alone.
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.
Recommended Free Tools
