How capable AI could become cheaper while frontier plans stay expensive

OpenAI now sells a $500-a-month personal plan. For a British customer, that sticker can plausibly feel like roughly £450 after currency conversion and VAT. That is not a mass-consumption price. It is the price of a professional capacity tier aimed at people who can turn greater usage and lower latency into work worth considerably more than the subscription.[1][2][4][5]

The useful question is not whether today’s exact Pro 500 service will suddenly be sold for the price of broadband. It is what must happen before a future service with comparable practical value becomes affordable to ordinary households. That requires separating the expensive frontier from the capability that eventually diffuses into cheaper models and plans.

A fixed capability level can become much cheaper

Stanford’s 2025 AI Index describes a fall from $20 to $0.07 per million tokens for models reaching roughly GPT-3.5-level performance on MMLU, using November 2022 and October 2024 endpoints. That is a comparison at a specified benchmark level, not a price forecast for whatever model leads the frontier next year.[8]

The workload mix changes the arithmetic

Here is an invented serving-cost example, not an API quotation. Suppose a service handles 500 monthly tasks. Assume a heavy route costs £0.20 per task and a light route £0.005. Sending all 500 through the heavy route costs £100. Sending 25 heavy tasks and 475 light tasks costs £5 + £2.375 = £7.375, about £7.38.

For an invented month of 500 tasks, sending every task to a heavy model at an assumed £0.20 each would cost £100. A mixed service sending 25 tasks there and 475 to a light model at an assumed £0.005 each would cost £5 + £2.375 = £7.375, or £7.38 rounded. These are original arithmetic assumptions, not provider margins, current tariffs or measured serving costs. A subscription also has costs beyond serving these tasks.

The saving depends on the light route completing those 475 tasks to the required standard. Misrouting difficult work creates retries and review cost. The mechanism is nevertheless straightforward: occasional access to an expensive route is different from a promise to use it continuously.

Capability can diffuse even while the frontier remains expensive

A future £35 plan does not need to reproduce every feature of today’s £455-equivalent tier to transform ordinary life. It needs to complete the tasks most people value: reliable research, file analysis, planning, tutoring, drafting and modest agent work. As those capabilities move into smaller and cheaper systems, the highest tier can keep selling the newest reasoning, greatest allowance and lowest latency to professionals.

A household-priced service and an expensive frontier tier can therefore coexist. The former sells enough capability for common jobs within bounded allowances. The latter can sell more of the scarce features: maximum effort, long-running work, priority service and greater usage. A fall in the cost of one fixed capability does not require the newest frontier tier to stop advancing.

Report an error or suggest a correction