The software we build with AI today will cost more to produce if we wait. I expect AI token costs to rise over the next year as major providers shift from growth-first pricing toward the margins they need to sustain their businesses. That is the direction I have been hearing from the industry, and it aligns with my analysis of the economics.

When a production input is getting more expensive, spending productively on it today can be cheaper than buying the same work later. The opportunity is to turn today's token spend into completed, tested software before the price of that development input climbs.

Development cost as token prices rise

For example, a software product costing $1 million to produce today would cost about $2.43 million to produce in three years if AI token prices rise by 50% annually. In this cost model, $600,000 of today's budget is tokens and $400,000 is other development expense. The scope, token quantity, and non-token costs stay constant, so the increase comes from paying more for the same AI-assisted development.

At that rate, the token portion grows from $600,000 to $2.025 million over three years. Add the unchanged $400,000 of other development expense and the same product costs $2.425 million, or about $2.43 million. Building now avoids a modeled $1.425 million increase in production cost, before counting the value of shipping earlier.

Development cost as token prices rise

Software development cost by years waited under three annual token-price increases A software product costs one million dollars to produce today, including 600,000 dollars in token expense and 400,000 dollars in other development cost. After three years, total production cost reaches 1.57 million dollars at 25 percent annual token-price increases, 2.43 million dollars at 50 percent, and 3.62 million dollars at 75 percent. $0M$1M$2M$3M$4M Today1 year2 years3 years Years waited before producing the same software Total development cost $1.57M$2.43M$3.62M 25% / year 50% / year 75% / year $1.00M
A software product costing $1,000,000 to produce today reaches $1.57M, $2.43M, or $3.62M in three years at annual token price increases of 25%, 50%, or 75%, respectively. The model assumes $600,000 of today's cost is tokens and $400,000 is other development expense, with scope and token volume held constant.

At the 50% annual increase, the same work costs $1.425 million more after three years because the token portion compounds while other development expense stays flat. Higher token shares or faster price increases create an even larger cost of waiting.

The opportunity cost is larger than the price difference

A two-year delay also means two years without the product in customers' hands. That can mean foregone sales, slower feedback, and lost time learning which features deserve investment. Those benefits belong in the decision. The build-versus-wait comparison weighs today's total production cost and earlier learning against higher future production cost and the value lost while waiting.

This is why I believe development budgets should account for rising AI costs now. Holding back a productive AI workflow to save today's tokens is false economy when the same development work will require more expensive tokens later. That is before counting the value of shipping earlier.

Spend for useful progress, not for token volume

There is an important distinction between not arbitrarily rationing tokens and consuming tokens without discipline. More tokens do not automatically mean more finished software. A model can loop, misunderstand the task, produce code that fails tests, or spend heavily on work that no customer needs.

The rational goal is to maximize accepted, useful product progress while the input is cheaper, not to maximize token usage. If a token budget blocks work that is producing reviewed, tested, valuable software, the budget may be optimizing the wrong thing. Keep controls for security, quality, scope, and stalled work. But set the spending threshold by the value of progress and the cost of alternatives, not by a desire to minimize the token line item at any price.

My position: AI token costs are going up, and the software we build with them will cost more to produce if we wait. Do not impose an arbitrary token cap that slows useful, verified progress. Spend to build the product now, while still holding the work to scope, tests, security, and customer value. The point is not to burn tokens; it is to avoid paying more for the same software later.

Spend decisively, with product discipline

  • Fund work that advances a validated product, not speculative scope with no customer or business case.
  • Keep quality gates: require working tests, review, security checks, and integration before counting generated code as delivered software.
  • Measure useful outcomes and cost per accepted feature. Redirect spend when agents stall, repeat work, or produce output that does not meet the bar.
  • Build so the product can take advantage of better models and tooling as they arrive, instead of hard-coding a costly dependency.

Research is a reason to measure actual task outcomes, not just count generated code or trust a productivity story. METR's early-2025 randomized study found that the experienced open-source developers in its sample took longer on assigned tasks with the AI tools they tested. METR's later update said its newer data was difficult to interpret because of selection effects and agent-based workflows. Neither result settles what will happen to a particular product team, but both argue against treating token consumption itself as proof of progress.

And when discussing financial statements, software development can be capitalized and amortized under certain rules, depending on the software's purpose and the applicable accounting standard. That accounting allocation is different from the economic question here: how much money and time will it take to produce the same useful product, and what is lost by delaying it?

Sources and notes

Model note: The chart uses three annual token-price increases (25%, 50%, and 75%) to quantify the cost of the upward trend I expect. These are planning rates, not published provider price announcements. The $1 million example allocates $600,000 to tokens and $400,000 to other development expense. The model holds scope, token quantity, and non-token costs constant. It covers software production only, not post-launch AI operating costs.

My practical takeaway is direct: software built with AI will cost more as token prices rise. Do not let an arbitrary token ceiling delay useful software. Build quickly, measure what the spending produces, and keep investing where the next unit of effort earns its keep.