AI cost fell more than 280-fold, Stanford shows

The cost of AI, measured by the price of each query to a language model, fell more than 280-fold in less than two years. Stanford University’s AI Index 2025 calculated, with data from Epoch AI and Artificial Analysis, that the price of using a model with GPT-3.5’s performance on the MMLU test (64.8 points) went from US$ 20 per million tokens in November 2022 to US$ 0.07 in October 2024, as a weighted average of three input tokens to one output token (Stanford HAI, 2025). That number measures the list price of a fixed capability, and company bills did not fall in the same proportion. In McKinsey’s 2026 global survey, one in five respondents says their organization already limits AI use because of operating costs, and Gartner predicts that inference cost per agentic workflow will more than quintuple through 2028. Tokens got cheap, companies started using far more of them, on tasks that call for more expensive models, and the largest cost of a project still sits outside the model bill: organized data, integration, human review and governance.

At Iasbeck&Co we implement CRM and AI agents in B2B operations, and the cost question comes up in almost every budget approval. I wrote this piece for the executive team and finance, who receive these numbers in presentations without a source, and for sales, marketing and CS leaders, who answer for the agents’ consumption. Measuring return and the full list of project costs are in AI ROI; here I stick to price, consumption, energy and cost per interaction.

How much has AI cost fallen, according to Stanford?

The figure is in chapter 1 of the AI Index 2025, on research and development, in the section on inference cost (pages 63 and 64). The method fixes a performance level on a test and looks, each month, for the cheapest model that reaches that level, so models of different capabilities are not mixed. On the GPQA test, PhD-level science questions, the price of scoring above 50% fell from US$ 15 per million tokens in May 2024 to US$ 0.12 in December. The report attributes the drop to increasingly capable small models and adds that hardware costs fall 30% a year while energy efficiency improves 40% a year.

The same report carries the caveat that usually stays out of the slides. Frontier models remain more expensive than smaller alternatives, and providers charge more for the top of the line than for small models from the same company. Comparing a reasoning model with the general model from the same provider, the AI Index found a cost nearly six times higher and a response 30 times slower. The 280-fold drop applies to anyone who accepts 2022 performance: ticket triage may work well at that level, while an agent that plans steps and queries several systems usually needs a more expensive model.

What does Epoch AI show about usage price and training cost?

Epoch AI, which supplies part of the AI Index data, published in March 2025 the analysis behind Stanford’s number (Cottier, Snodin, Owen and Adamczewski, Epoch AI, 2025). The authors tracked models on six tests over three years and measured how fast the price of reaching each performance milestone fell. The decline ranged from 9 to 900 times per year depending on the milestone, and the price of reaching GPT-4’s performance on the science questions fell 40 times per year. Epoch itself warns that the fastest drops happened in the last year of the series and that it is unclear whether they will continue.

On the side of those who build the models, the curve goes the other way. In June 2024, Epoch estimated that the amortized hardware and energy cost of the final training runs of frontier models has grown 2.4 times per year since 2016, and that the largest runs would exceed US$ 1 billion by 2027 if the trend continued (Cottier, Rahman, Fattorini, Maslej and Owen, Epoch AI, 2024). In the breakdown for models such as GPT-4, hardware accounts for 47% to 67% of development cost, research staff for 29% to 49% and energy for 2% to 6%. Buyers pay for that training through the price of frontier models and plans, which is why the most capable model in the catalog stays expensive while the one from two years ago costs cents.

Which studies measure the cost of AI?

Source	Year	Sample or method	Main finding
Stanford HAI, AI Index 2025	2025	API prices (Epoch AI and Artificial Analysis) for a fixed performance level	Price of GPT-3.5 level fell from US$ 20 to US$ 0.07 per million tokens between Nov/2022 and Oct/2024
Epoch AI, inference prices	2025	Six performance tests over three years, lowest price to reach each milestone	Decline of 9 to 900 times per year, depending on the milestone
Epoch AI, training cost	2024	Estimate of hardware and energy for frontier models	Cost grows 2.4 times per year since 2016; energy is 2% to 6% of the total
Gartner, inference forecast	2026	Analyst forecast for AI providers	Inference will cost providers over 90% less in 2030; agents use 5 to 30 times more tokens per task
Gartner, inference paradox	2026	Analyst forecast for product leaders	Inference cost per agentic workflow more than quintuples through 2028
McKinsey, The state of AI	2026	Online survey of 1,719 participants in 97 countries, May to June 2026	One in five says their organization limits AI use because of operating costs
International Energy Agency, Energy and AI	2025	Modeling with electricity consumption scenarios	Data centres used 415 TWh in 2024 and should reach about 945 TWh in 2030

Why does the AI bill rise if tokens get cheaper?

Gartner published two forecasts in 2026 that, read together, describe the movement. In March, the firm predicted that running a one-trillion-parameter model will cost providers over 90% less in 2030 than in 2025 (Gartner, 2026). The same release warns that this decline will not be fully passed on to enterprise customers and that agentic models require 5 to 30 times more tokens per task than a standard chatbot. Since consumption grows faster than prices fall, Gartner concludes that overall inference costs should rise.

In August, Gartner named the phenomenon the “inference paradox” and predicted that inference cost per agentic workflow will more than quintuple through 2028 (Gartner, 2026). According to the release, routing a task to an agentic reasoning model raises the provider’s inference cost by at least five times compared with a basic chatbot interaction. Both forecasts were written for companies that sell AI, and buyers receive that cost through the price of credits and plans.

McKinsey’s survey shows the effect on the buyer’s side: one in five respondents says their organization limits AI use because of operating costs, including tokens, and about one in ten reports that limit for each of three tool types, chatbots, agents and coding agents (McKinsey, 2026). The full reading of that survey is in the state of AI in companies. In Brazil, cost weighs before use: in TIC Empresas 2025, 42% of large companies that did not use AI said costs seem too high (Cetic.br/NIC.br, 2026), as I discuss in AI in Brazilian companies.

How much does data centre energy weigh?

The International Energy Agency’s Energy and AI report estimates that data centres consumed about 415 TWh in 2024, around 1.5% of the world’s electricity, with 45% of that consumption in the United States, 25% in China and 15% in Europe (IEA, 2025). Consumption has grown about 12% a year since 2017, more than four times the pace of total electricity consumption. In the base case, it more than doubles by 2030 to about 945 TWh, slightly above Japan’s current consumption, with AI as the main driver.

The AI Index 2026 records that AI data centre power capacity reached about 29.6 GW in the fourth quarter of 2025 (Stanford HAI, 2026). In a study cited by the report, energy per medium-length query ranged from about 5 to 23 Wh among the 15 most energy-hungry models of 2025, and the choice of model weighs on energy as much as on the bill.

For the budget of a commercial team, energy comes in through the provider’s price, with no separate line on the invoice. It limits how far frontier model prices can fall and enters the sustainability report of companies that disclose indirect emissions.

Where is the cost of an AI project?

The token price is the only item that follows the Stanford and Epoch curves, and it only falls for the company when the platform passes the drop on. In the projects I follow, model consumption is almost never the largest line in the first year, and data preparation and review hours usually weigh more.

Component	Does it follow the token price drop?	What drives the cost
Model consumption per task	Partly	Token price, context size, number of calls and choice of model
Platform credits	Depends on the vendor	Price per action set in the contract, which may not change when tokens get cheaper
Data preparation	No	Standardized fields, duplicates, reliable history in the CRM
Systems integration	No	Number of systems the agent reads and where it writes
Human review	No	Response quality and approval rule set by leadership
Governance and security	No	Usage policy, access control, sample audits
Team training	No	How many people operate, review and configure the agents

How do you estimate the cost of an AI agent per interaction?

The per-interaction math has two parts. The first is model consumption: input tokens times the input price, plus output tokens times the output price, multiplied by the number of calls the agent makes to complete the task. The second is reviewer time, which platforms do not show. On platforms that charge credits per action, replace the first part with the average credit consumption measured in a test, converted into local currency, as I describe in AI agents.

Illustrative example: customer service agent

The numbers below are hypothetical and do not represent any vendor’s price or any client case. Picture an agent that classifies each ticket, looks up the answer in the knowledge base, drafts a reply and checks its own text, in four calls to the model. Each call reads about 6,000 tokens (instructions, customer history and knowledge base excerpts) and generates about 800. At a hypothetical price of US$ 1 per million input tokens and US$ 5 per million output tokens, with an exchange rate of R$ 5.50 just for the math, consumption comes to US$ 0.04 per ticket, or about R$ 0.22. The analyst takes two minutes to review each draft, at a loaded cost of R$ 60 an hour, which gives R$ 2.00 per reviewed ticket. The operation receives 5,000 tickets a month.

Scenario (5,000 tickets a month)	Model per month	Review per month	Total per month	Total per ticket
A. Mid-tier model, 100% review	R$ 1,100	R$ 10,000	R$ 11,100	R$ 2.22
B. Mid-tier model, 20% sample review	R$ 1,100	R$ 2,000	R$ 3,100	R$ 0.62
C. Reasoning model (5 times the cost), 20% sample	R$ 5,500	R$ 2,000	R$ 7,500	R$ 1.50
D. Tokens 90% cheaper, 100% review	R$ 110	R$ 10,000	R$ 10,110	R$ 2.02

Scenario D shows why the token price drop barely reaches the bill. A 90% drop in token price, the size Gartner forecasts for providers through 2030, takes 9% off the monthly bill when review is total. Moving from total review to a 20% sample, which depends on response quality and a leadership decision, cuts 72%. Scenario C shows the opposite risk: a reasoning model, at the minimum multiplier Gartner cites, quintuples consumption and overtakes review as the largest line in the bill.

Sales agents tend to read more context: one that researches the account before a meeting reads the website, news and CRM history, and input per task can be several times that of the example. The estimate is only valid after a test with a few dozen real cases.

What these data do not say

Stanford’s number measures API list price, and price differs from cost: providers can subsidize models to gain market share or charge high margins at the top of the line. The average of three input tokens to one output token does not represent an agent that reads 6,000 tokens and writes 800, and MMLU is a multiple-choice test that does not measure whether the model solves your task, with your data and in your language. The report also describes the drop as happening in about 18 months, but the dates it gives add up to 23; the order of magnitude does not change, and I prefer to cite the dates.

Epoch’s rates cover windows of one to three years, and Epoch itself doubts the fastest drops will repeat. Gartner’s forecasts are analyst estimates, with no sample published in the release. McKinsey relies on self-reporting, and the IEA scenarios open a range of 700 to 1,700 TWh for 2035.

None of the sources measures the total cost of an AI project at a user company, and I found no primary research that does so with a Brazilian sample, nor a public series of token prices in Brazilian reais. A warning that companies miss their AI cost estimates by 500% to 1,000% also circulates in news articles, attributed to Gartner analysts at a 2024 keynote. I found no public release or report from the firm with that number, so it stays out of this analysis.

What leadership does with these numbers

1. Separate price from consumption. Ask for the price per action and measure consumption per task in a test, because a token price drop only helps if consumption stays stable.
2. Choose the model by task. Classification, extraction and summarization usually work with smaller models, and the reasoning model is for multistep tasks with decisions, as Gartner itself recommends.
3. Calculate cost per interaction with review included. Take finance the total per ticket, proposal or contact, using your operation’s numbers.
4. Set a consumption cap per agent. Each agent gets a monthly cap and an owner who receives the alert.
5. Plan the reduction in review. Write the criterion, with correction rate and deadline, that allows moving from total review to sampling, because that decision moves the bill more than the model price.
6. Put the pass-through in the contract. Negotiate a review of credit prices when model prices fall and track consumption by process.
7. Review cost and results in the same meeting. Bring consumption and cost per interaction to the review where the team presents its business indicator, as I describe in AI ROI.

AI cost and training: what HR and L&D need to know

Cost per interaction depends on behavior. Someone who pastes a whole document into the assistant when an excerpt would do multiplies the input tokens, and a leader who does not trust the drafts keeps total review for months and pays the most expensive line in the example. HR recognizes the need for training when credit consumption grows without more volume and when nobody on the team knows what an agent action costs. Leaders learn to read cost per interaction and to decide when to reduce review, and those who configure agents learn to trim context and choose the model per step. To evaluate, compare consumption per task and correction rate in the four weeks before and after the training.

Model prices will keep falling, and the part of the bill the company controls lies in agent design, in the data agents read and in the team’s confidence to review less. That is the work of the in-company applied AI training, with groups for the executive team, leaders and the people who operate the agents. It delivers the prompt library by area and the agent architecture, with what each agent reads, what it executes on its own and who approves the rest, the basis for estimating and controlling the cost of each interaction.

Frequently asked questions

How much has the cost of AI fallen in recent years?

According to Stanford’s AI Index 2025, the price of using a model with GPT-3.5-level performance fell from US$ 20 to US$ 0.07 per million tokens between November 2022 and October 2024, more than 280-fold. Epoch AI found declines of 9 to 900 times per year, depending on the performance level. Frontier models remain much more expensive than smaller ones.

What is cost per token?

A token is the unit of text the model processes, and Gartner uses a reference of about four characters per token. Providers charge per million tokens, with different prices for what the model reads (input) and what it writes (output), and output usually costs more.

Why does a company’s AI bill rise if the token price falls?

Consumption grows faster than prices fall. Gartner estimates that agents use 5 to 30 times more tokens per task than a standard chatbot and forecasts that cost per agentic workflow will more than quintuple through 2028. In McKinsey’s survey, one in five respondents says their organization already limits use because of cost.

How much does an AI agent cost per interaction?

It depends on the model, the context the agent reads, the number of calls per task and human review. In this article’s illustrative example, model consumption comes to cents per ticket and total human review costs about nine times more. Measure tokens or credits per task in a test with real cases before projecting volume.

Will data centre energy make AI more expensive?

The IEA estimates that data centre consumption will more than double by 2030, to about 945 TWh, with AI as the main driver. Energy accounts for 2% to 6% of the development cost of frontier models, according to Epoch, and reaches companies through provider prices, with no line of its own on the invoice.

How do you prepare the team to use AI without blowing the budget?

Teach leaders to calculate cost per interaction and to decide when review can move to sampling. Those who configure agents need to learn to trim context and choose the model per step. Track consumption per task and correction rate before and after the training.