In November 2021, using the only model that could reach a certain level on a standard AI benchmark cost $60 per million tokens. Three years later, a model at the same level cost $0.06. Few prices in the history of technology have fallen that fast.
So what will AI cost in 2030, or in 2036? Nobody can give an exact number, but the forces that set AI prices are well measured now, and they point in a clear direction, with some important exceptions.
The short version: The price of any given level of AI capability has been falling roughly 10 to 13 times per year, and it's likely to keep falling over the next decade, though more slowly. Everyday AI (chat, summaries, simple automation) is heading toward almost free. The newest, most capable models will keep a premium price, and total AI spending will probably grow, because cheaper AI gets used far more. Electricity and chips will set the floor.
What has happened to AI prices so far?
Three independent analyses tell the same story:
- a16z (November 2024) found that, for an AI model of equivalent performance, the cost fell about 10 times every year: from $60 per million tokens for GPT-3 in 2021 to $0.06 for a much smaller model reaching the same score, a 1,000-fold drop in three years.
- Stanford's AI Index 2025 reported that running a system at the level of GPT-3.5 fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a more than 280-fold reduction.
- Epoch AI (September 2026), in the most detailed study so far, estimated that the cost of reaching a given level of AI performance has fallen about 47% per quarter since 2023, or 13 times per year. In one example, a level of performance on hard science questions that cost about 30 cents per question in early 2025 cost a tiny fraction of a cent less than 18 months later: a 725-fold drop.
One detail matters for any forecast: these numbers track the price of a fixed level of capability, usually reached later by a smaller, cheaper model. They don't mean the newest model gets cheaper every month.
Why AI keeps getting cheaper
Several forces push in the same direction at once:
- Better chips. Stanford's AI Index found that the cost of AI hardware has been falling about 30% a year, while energy efficiency has improved about 40% a year.
- Smarter software. Engineers keep finding ways to get the same results with less computing power, through better training, compression and serving techniques.
- Smaller models catching up. Capabilities that once needed the largest model can be reached a year or two later by a model small enough to run cheaply, sometimes even on a phone or laptop.
- Competition and open models. Many companies now sell similar capabilities, and openly available models keep the pressure on prices.
Epoch AI also found that prices fall fastest right after a level of performance first appears. When a new top model arrives, its maker can charge a premium for a while; once competitors catch up, the price for that level drops quickly.
Why your AI bill may not shrink
If AI is getting so much cheaper, why do companies keep spending more on it? Three reasons.
1. The best models stay expensive
In late 2024, a16z noted that OpenAI's leading model at the time cost the same per output token as GPT-3 had at launch: $60 per million. Providers concede the low end of the market and charge a premium for the newest capabilities. The same pattern shows up in subscriptions: alongside the familiar $20-a-month plans, the major AI companies added premium plans at around $100 to $200 a month for their most capable models and heaviest users.
2. Cheaper AI gets used much more
When something becomes cheaper, people find many more uses for it. AI agents that plan, check their work and call other tools can use far more tokens to finish one task than a simple chat answer. So even as the price per token falls, the number of tokens a business uses tends to rise, and the total bill can grow.
3. Electricity and data centres
AI runs on data centres, and data centres run on electricity. The International Energy Agency estimates that data centres used about 415 terawatt-hours of electricity in 2024, around 1.5% of the world's total. In its base case, that more than doubles to about 945 TWh by 2030, with AI as the most important driver, and rises to around 1,200 TWh by 2035. Electricity used by the servers that run AI is projected to grow about 30% a year. Power, chips and new data centres take years to build, and their cost is the floor that AI prices can't go below.
Our forecast: AI prices from 2026 to 2036
What follows is our reading of the trends above, not a guarantee. The further out the forecast goes, the wider the range of outcomes.
2026–2028: Everyday AI becomes almost free
- The price of a given capability keeps falling fast, likely several times per year.
- Chat, summaries, translation, classification and simple automation cost so little that they become a standard feature of ordinary software, not a separate purchase.
- The newest models still carry a premium, and premium subscription tiers stay.
- Billing shifts from "per seat" to "per use", especially for AI agents.
2029–2032: The decline slows, and AI moves onto devices
- Price drops continue but slow down as the easy efficiency gains are used up and electricity, chips and data-centre capacity become the main constraints.
- More AI runs directly on phones, laptops and office servers, where each extra question costs nothing.
- More companies price AI by the result (a resolved support ticket, a processed invoice, a qualified lead) instead of by the token.
2033–2036: Today's best AI becomes a commodity
- Even if the yearly decline slows to 2 or 3 times, ten years of compounding is a 1,000-fold to 59,000-fold drop. Capabilities that need today's most expensive models should cost a tiny fraction of today's price.
- Energy becomes the biggest single part of the cost of running AI.
- There will still be a premium tier: whatever the newest, most capable systems are in 2036 will cost more, just as they do today.
The likely shape of the market: a very cheap, almost free layer for everyday tasks, and a smaller, expensive layer for the most advanced work.
What this means for your business
- Don't pay frontier prices for simple tasks. Use the cheapest model that does each job well, and keep the expensive models for the work that needs them.
- Build so you can switch models. Prices and leaders change every few months. Standards like the Model Context Protocol make it easier to swap the AI behind your systems.
- Budget per task, not per tool. Know what one answered customer, one processed invoice or one report costs you, and watch that number over time.
- Don't wait for AI to get cheap. It already is for most business tasks. The value comes from your data and processes being ready to use it, which takes time to build. Our guide to AI agents for business shows where to start.
- Expect software itself to cost less. AI tools make building websites and systems faster, and those savings should reach clients too. That's why we've lowered our prices for 2026.
Frequently asked questions
Will AI get cheaper in the next 10 years?
Very likely, for any given level of capability. Since 2023, that price has fallen about 13 times per year according to Epoch AI. The pace will probably slow, but even a much slower decline compounds to a huge drop over ten years. The newest, most capable models will keep a premium.
Will ChatGPT and other AI subscriptions get cheaper?
Standard plans have stayed around $20 a month while offering much more for the money, and premium plans for heavy users now cost around $100 to $200 a month. Expect prices to stay roughly stable while what you get keeps growing, with more free features and more pay-per-use options.
Will electricity make AI more expensive?
Electricity sets the floor. The International Energy Agency expects data-centre electricity use to more than double by 2030, driven mainly by AI. Efficiency keeps improving, so the cost per answer should still fall, but power supply could slow how fast prices drop and how quickly new capacity comes online.
Should my business wait for AI to get cheaper before using it?
No. For most business tasks, such as answering customers, summarizing documents and preparing reports, AI is already inexpensive. The bigger cost is getting your data and processes ready, and that work only gets more valuable as AI gets cheaper.
Where to go from here
The winners of the next decade won't be the companies that waited for AI to become cheap. They'll be the ones whose systems were ready to use it when it did.
At GTech, we build custom software, CRMs and ERPs that are ready for AI, and we help you choose the right model for each task so you don't overpay. Tell us what you'd like to automate.
Sources
- Epoch AI: "The plunging price of thought" (September 2026)
- Epoch AI: "LLM inference prices have fallen rapidly but unequally across tasks" (March 2025)
- a16z: "Welcome to LLMflation – LLM inference cost is going down fast" (November 2024)
- Stanford HAI: AI Index Report 2025
- International Energy Agency: Energy and AI (2025)
