It’s getting cheaper to own good AI, as opposed to renting it by the token.
The meter shapes how we work
Whether we pay by the token for every piece of text an AI reads and writes, or by a monthly subscription with a limit on usage, there’s a meter running.
A meter trims your question, skips your second attempt, and holds the long job back until the limit resets. Nobody bargains over these concessions out loud. It just happens.
Also, the meter comes with terms.
Whoever runs the meter sets the terms
Three dated examples.
Cursor, a coding tool now owned by SpaceX, which merged with xAI, changed its pricing on 16 June 2025. In a post on 4 July 2025 it wrote: “We recognize that we didn’t handle this pricing rollout well, and we’re sorry.” It offered “usage refunds for those with unexpected usage between June 16 and July 4.”
Anthropic, the company behind the Claude models, added weekly limits to its paid plans in August 2025, TechCrunch reported on 28 July. Its help page today reads: “Both Pro and Max plans have a five-hour session limit and a weekly limit.”
OpenAI announced on 29 January 2026 that on 13 February it would retire GPT‑4o, GPT‑4.1, GPT‑4.1 mini and OpenAI o4-mini from ChatGPT. Its stated reason was low use, “with only 0.1% of users still choosing GPT‑4o each day.”
Cursor apologised and offered refunds. OpenAI gave its numbers. Fair enough. Still, the terms changed in all three cases, whether the company meant them to or not. The terms of your metered AI are set by whoever runs the servers, not by you.
The same answer costs a fraction of what it did two years ago
According to Stanford’s AI Index 2025, the cost of running AI at a fixed level of quality, that of GPT-3.5, fell more than 280-fold between November 2022 and October 2024.
Epoch AI, a research group that tracks AI trends, looked at the price of reaching the same score on a test over time. In March 2025 it reported: “the rate of decline varies dramatically depending on the performance milestone, ranging from 9x to 900x per year.” It added its own caution: “The fastest price drops in that range have occurred in the past year, so it’s less clear that those will persist.”
The hardware underneath moved too. The AI Index again: “At the hardware level, costs have declined by 30% annually, while energy efficiency has improved by 40% each year.”
Lower prices are great for renting. They also mean good AI can now run on a machine you buy yourself.
In August 2025, one gaming graphics card ran the best AI of a year earlier
In August 2025 Epoch AI compared the best AI of each month with what a single gaming graphics card can run. Its finding: one NVIDIA RTX 5090, “under $2500”, runs open models you can download that match the best AI from 6 to 12 months earlier.
Under $2,500. That’s a card for a gaming PC. Not a laptop, and not a data centre server either.
For the same amount of work, models have also got smaller. Researchers at Tsinghua University measured open models from February 2023 to April 2025: the most ability packed into a model of a given size doubled about every 3.5 months. The same work can be done by smaller models, on less powerful machines.
On top of that, hardware costs fall 30% a year.
These trends have end dates. They cover the last two to three years. We don’t know how prices will behave from here, and Epoch’s own caution on its price figures applies.
Requests got longer, and one provider’s token count grew by 6 billion a minute in a quarter
Four measurements, each with its own limits.
OpenRouter, a service that passes AI requests on to many providers, studied “over 100 trillion tokens” of its traffic and published the results with the investment firm a16z in December 2025. The average prompt went up roughly fourfold, from around 1.5K tokens in early 2024 to over 6K. Models that reason step by step before answering went from a negligible share of usage in early 2025 to more than half of all tokens by the end of the year. That is one service’s customers, not all AI use.
METR, a research group, measures how long a task the best AI of the day can finish, counted in the time a human typically needs for tasks the AI completes with a 50% success rate. For the best models, that “time horizon has doubled approximately every seven months since 2019” (February 2026 version). Its tasks are research and software work, not everything people do.
On 22 July 2026 Alphabet told investors that the AI it sells to other companies’ software processed about 22 billion tokens a minute, up from 16 billion one quarter earlier. One provider, one quarter.
These four sources speak as of their own dates.
Here the two ways of paying split. A meter bills you per token. A machine you own costs you its purchase price and the power it uses, whatever you run on it.
The case for renting: cheaper per token, and in March 2026 the best AI was closed
Critics of owning make two points.
First, price. By spreading many customers across their chips, large providers can rent you AI at a lower price per token than you’d pay to buy and power a machine running a downloaded open model. Epoch’s figures above show how fast those rental prices fell up to early 2025.
Second, quality. Stanford’s AI Index 2026 uses the Arena leaderboard, a public ranking where people judge answers without knowing which company’s model wrote them. In March 2026 the best closed model led the best open model there by 3.4%. In August 2024 the gap was 0.5%.
We have no numbers of our own to set against either point, and in March 2026 the top AI on that leaderboard was one you rent.
But price per token was never the economics of owning. For many tasks, a card under $2,500 that runs open models matching the best AI of 6 to 12 months before is enough. The real question is: who sets the terms for you?
The quality gap between open and closed models moves in both directions: 15.2% in May 2023, 0.5% in August 2024, 3.4% in March 2026.
Owning one AI and renting another work side by side
That doesn’t mean giving up the best rented AI for the hardest tasks. Rent it when you need it. Your own AI sits next to it, on terms you choose.
We build Anthill, which runs AI on your own machine, and we document providers in the AI Ownership Index.
Christoph Welsbach, founding director of OneHill. 24 October 2026.
Written by Christoph. His own AI was used to help edit and structure this piece.
Sources
| Claim | Source URL | Date checked |
|---|---|---|
| Cursor pricing change 16 June 2025; apology and refunds 16 June to 4 July, posted 4 July 2025 (verbatim quotes) | https://cursor.com/blog/june-2025-pricing | 2026-10-22 |
| Cursor’s maker Anysphere bought by SpaceX for $60B, deal closed August 2026; xAI merged with SpaceX in 2026 | https://www.vktr.com/ai-market/spacex-completes-60-billion-cursor-acquisition/ ; https://www.techzine.eu/news/devops/142197/spacex-acquires-cursor-for-60-billion/ | 2026-10-22 |
| Anthropic weekly limits on paid plans, reported 28 July 2025, effective 28 August 2025 (secondary, TechCrunch) | https://techcrunch.com/2025/07/28/anthropic-unveils-new-rate-limits-to-curb-claude-code-power-users/ | 2026-10-22 |
| ”Both Pro and Max plans have a five-hour session limit and a weekly limit.” (provider’s own help page) | https://support.claude.com/en/articles/11145838-use-claude-code-with-your-pro-or-max-plan | 2026-10-22 |
| OpenAI retires GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini from ChatGPT on 13 Feb 2026; “with only 0.1% of users still choosing GPT‑4o each day”; posted 29 Jan 2026 | https://openai.com/index/retiring-gpt-4o-and-older-models | 2026-10-22 |
| Cost of a GPT-3.5-level system fell over 280-fold, Nov 2022 to Oct 2024; hardware costs down 30% a year, energy efficiency up 40% a year (AI Index 2025, Top Takeaways, p. 4, item 7) | https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf | 2026-10-22 |
| Price decline 9x to 900x a year by performance milestone; caution on persistence (Epoch AI, 12 Mar 2025) | https://epoch.ai/data-insights/llm-inference-price-trends | 2026-10-22 |
| One RTX 5090 (under $2,500) runs open models matching the frontier of 6 to 12 months earlier (Epoch AI, 15 Aug 2025) | https://epoch.ai/data-insights/consumer-gpu-model-gap | 2026-10-22 |
| Maximum capability density of open models doubles about every 3.5 months, Feb 2023 to Apr 2025 (Xiao et al., Nature Machine Intelligence, 6 Nov 2025) | https://www.nature.com/articles/s42256-025-01137-0 ; https://www.tsinghua.edu.cn/en/info/1245/14611.htm | 2026-10-22 |
| Prompt tokens per request about 1.5K to over 6K since early 2024; reasoning-model share over 50% in 2025; over 100 trillion tokens; OpenRouter with a16z, Dec 2025 | https://openrouter.ai/state-of-ai | 2026-10-22 |
| 50% time horizon definition; doubled about every seven months since 2019 (METR, arXiv 2503.14499 v3, 25 Feb 2026) | https://arxiv.org/html/2503.14499v3 | 2026-10-22 |
| About 22 billion tokens per minute through Alphabet’s model APIs, up from 16 billion a quarter earlier (Alphabet Q2 2026, 22 Jul 2026) | https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/ | 2026-10-22 |
| Arena gap top closed vs top open: 15.2% (May 2023), 0.5% (Aug 2024), 3.4% (Mar 2026); blind, community-driven ranking (AI Index 2026, ch. 2, p. 77) | https://hai.stanford.edu/assets/files/ai_index_report_2026_chapter_2_technical.pdf | 2026-10-22 |