The cost of artificial intelligence has dropped thousands of times in recent years—faster than the price of any other revolutionary technology in the past century. Analysts at Epoch AI estimate that the cost of artificial intelligence will drop by nearly 50% every quarter and 13 times for the full year. Lithium batteries, DNA sequencing technology, and even computing power, these technologies are developing rapidly in accordance with Moore’s Law, but their prices are not falling so fast.
Image credit: Roman Budnikov / unsplash.com
Reducing costs are driving technology adoption but are also raising difficult questions about whether to commit to cutting-edge artificial intelligence development; service providers are finding it increasingly difficult to cultivate customer loyalty and generate ongoing revenue. If a person knows that in just a few weeks or months the service he is interested in will become more available and cheaper, what is the point of implementing it immediately? Why continue with your current supplier if a competitor offers a better, more affordable solution? Industry leaders like OpenAI and Anthropic find themselves in a dilemma: They can hardly expect to make huge profits from cutting-edge models that outperform their competitors if the advantage is only short-term.
In just five years, the cost of AI has dropped hundreds of thousands of times. In comparison, the price of lithium batteries fell only a hundred times from 1991 to 2024. In the 60 years since 1940, computing equipment has become hundreds of billions of times cheaper. For much of this time, Moore’s Law has been at play: the productivity and efficiency of computing has improved dramatically every year—modern computers are not only faster than their predecessors, but also cheaper to run. DNA sequencing demonstrates the most similar price decline—the relative cost of the process fell even more than the cost of artificial intelligence, but it took just over 20 years. The cost of artificial intelligence is falling by nearly 50% quarterly and 13 times annually, which is 4 times faster than DNA sequencing and 18 times faster than lithium batteries.
The statistics presented come with a number of caveats. The data for computing equipment only goes back to 2001, and the statistics for electricity only go back to 1973, and does not take into account the explosive growth of solar energy in recent years. Tracking the value of artificial intelligence only begins in 2023, with earlier data based on current short-term trends and third-party research. Epoch AI’s artificial intelligence analysis is not based on any specific model, but on a set of models that show at least 81.25% on the graduate-level Google-Proof Q&A (GPQA) diamond benchmark – taking into account the cost of having the model answer a question with three possible options. This helps assess AI capabilities in different areas of knowledge, but does not ensure that the entire range of AI capabilities is covered. It’s clear that these tasks are easy for the new models – they score well in tests and cost less.
For example, the OpenAI o3 model released in January 2025 scored 75% in the GPQA diamond test and cost $0.30 per question. Just a year and a half later, OpenAI GPT-5.6 Luna was released, showing the same results in the same tests at a price of $0.0004 per question. The cost of AI is falling unevenly. In the GPQA Diamond benchmark, there are periods of sharp cost decline, followed by a stagnant phase and reaching a steady state – more stable dynamics found in the AIME OTIS math test and chess problems. There will be virtually no progress in higher mathematics between 2025 and 2026, with costs only falling sharply mid-year.
Image credit: Igor Omilaev / unsplash.com
In most areas, the pace of cost reductions is slowing: initially, prices fell by 66% per quarter across the five benchmarks, but two years later this figure fell to 32% per quarter. This means that either progress in AI is slowing down, or that reducing operating costs is becoming increasingly difficult—especially in 2026, as the price of electricity and computing equipment increases. If the resources required to operate artificial intelligence are more expensive than before, it will be difficult to reduce the price of services.
The main problem with statistical analysis is “benchmarking” – developers purposely train their models to perform well on tests designed to test general abilities. When conducting their tests, Epoch AI experts used an element of randomness that prevented the models from “understanding” that they were being tested, and the developers from being able to purposefully train them to successfully pass third-party testing. This approach was not applied to all tests – in its absence, higher cost reduction rates were observed, which indirectly confirms the existence of certain optimizations for a specific benchmark. This does not invalidate the results, but it does give reason to be skeptical of these results.
Developers of artificial intelligence, especially advanced artificial intelligence models, who are willing to spend hundreds of billions of dollars on infrastructure, worry about raising questions about the wisdom of high tariffs. Anthropic Mythos and Fable class models, as well as the latest OpenAI models, are more expensive to run than similar models, but the cost of processing tokens is steadily declining. Now, leading developers are competing on the Pareto frontier, the so-called sweet spot between cost efficiency and model “smartness” level. Is it worth paying extra for premium features if other models can deliver 90% of the leader levels at a fraction of the price? Especially with open scale models, third-party service providers can compete with each other by offering the most efficient and affordable versions of these systems.
Habits play a big role for customers: the interfaces, established workflows, and overall software ecosystem in the organization; the trust factor also plays a huge role, especially when it comes to long-term obligations that involve financial costs. Switching to a new model requires verifying the accuracy of stated performance data and ensuring future pricing does not radically change, thereby erasing any savings. It’s important to be confident that familiar systems won’t catch up to competitors in terms of features within a week, and that new systems don’t contain bugs or vulnerabilities that would compromise the confidentiality of materials. Therefore, the choice is not only limited by the price or smart capabilities of the model. Artificial intelligence is getting cheaper, but that doesn’t mean the subscription model is losing its right to exist — it’s just becoming harder to prove its viability.
If you find an error, select it with your mouse and press CTRL+ENTER.