Frontier models got cheap faster than anyone budgeted for

The price of a capable token fell by more than half again this quarter. When intelligence gets this cheap, the question stops being 'can we?' and becomes 'should we?'.

Roughly two years ago, running a strong model over every support ticket, every document, every code review would have been absurd on cost alone. This quarter the same work costs a fraction of that. The curve has bent faster than almost anyone's budget assumed.

What suddenly becomes affordable

  • Running analysis over everything, not a sampled slice. Read every ticket, not one in ten.
  • Multiple passes per task. Draft, critique and revise for the price of a single call a year ago.
  • Background work nobody used to justify: labelling, summarising and enriching data continuously rather than on demand.
When a thing gets ten times cheaper, you don't do the same amount for less. You do things you would never have started.

The catch

Cheap tokens make it easy to spend a lot of them badly. The teams getting value are the ones treating model calls like any other unit cost: measured, capped and attached to an outcome. The teams in trouble are the ones who noticed the price drop and stopped counting.

Worth your time

← All issues Get the next one by email →