The Cost of AI in the Coming Years: Cheaper Units, Heavier Bills?

When Epoch AI measures the cost of matching GPT-4 on a hard science test called GPQA, it finds that the price fell about 40-fold a year. Yet in September 2026 The New Stack described subscription plans in which the allowance is set per session and per week. How can cheapness and tight allowances coexist?

A cheaper unit

Epoch ran its analysis on state-of-the-art models across six tests and found that the annual fall in the price of reaching a given performance level ranged from 9-fold to 900-fold depending on the test. The researchers qualified this: the fastest declines occurred in the last year of the period analysed, and it is less clear that they will persist. The trend is real; mechanically extending its line is not guaranteed. For the debate over the size of the economic gain, see the article "The Economics of AI: Why Do Economists Disagree on the Gain?"

A bill that may grow heavier

Cheapness is measured per token; the bill is measured in the number of tokens. In describing a multi-agent research system, Anthropic reported that agents typically use about four times the tokens of a chat, and multi-agent systems about fifteen times. On the BrowseComp test it found that token usage alone explains 80% of the variance in performance. My inference, not a quotation, is that better performance on complex tasks is often bought with more tokens, so spending grows even as the unit gets cheaper.

On the commercial side, The New Stack reported that Anthropic imposes a weekly cap on top of the session cap, and that OpenAI's coding tools may share one allowance with other agentic tools, with extra credit available once it runs out. Limits are the providers' way of managing the capacity available.

A scarce supply of hardware

S&P Global said in January 2026 that AI demand for high-bandwidth memory (HBM) is drawing on DRAM manufacturing capacity, tightening the supply of conventional memory and raising its prices. This does not mean the cost of inference will necessarily rise, but it shows that the unit price is not the only factor.

What can be anticipated?

There is no documented source estimating that the cost of "effective intelligence" will multiply tenfold or more, so we leave such figures out. What can be said as analysis, not forecast: in the coming years, God willing, it is plausible that two tracks will sit side by side, one on which the cost of the usual level keeps falling and one on which heavier uses are rationed, and that the effect will vary by type of use. Knowledge of what has not happened rests with God alone, so the safest course is to build habits that hold in either case.

What to do today

These are operational measures, not financial or investment advice:

• Start with the smallest model adequate for the task, and use the strongest only where it makes a visible difference.
• Defer what does not need an immediate answer: both Anthropic's and OpenAI's APIs discount batch requests by 50%, and OpenAI gives a 24-hour completion window.
• Reuse fixed context: Anthropic prices reads from cached context as a fraction of the base input price, 10% for most models and lower for some.
• Do not launch an agent on a task a single question would settle.
• Record monthly what you spent and what you produced; what is not measured is hard to control.

For non-profit associations, a balancing view is in the article "Board the AI Train While the Tickets Are Free Today". And the old rule stands: take the means in planning and measuring, and rely on God for provision.