Two pricing events in nine days: Cursor's ₹649 India plan on 28 July, and OpenAI cutting GPT-5.6 Luna by eighty percent on 30 July.
If your team's AI budget is a per-seat number written down at the start of the year, it is now wrong in both directions — you are overpaying somewhere and under-provisioned somewhere else.
Per-seat is the wrong unit
Seat pricing made sense when the tool was an autocomplete. It stopped making sense when agents started running long jobs, because two developers on the same plan can differ by an order of magnitude in what they consume.
Cursor's Router, shipped 22 July, exists precisely because request cost varies enormously by task. Budget the same way it routes: by workload, not by headcount.
Three numbers worth tracking
- Cost per merged pull request — the only figure that connects spend to output
- Share of requests going to the cheap tier — if it is low, you are paying premium rates for classification work
- Cost of your most expensive single workflow, which is usually one automated job nobody has looked at since it shipped
Assume the price falls again
An eighty percent cut on one tier in one month is not a stable market. Do not sign long commitments on today's rates, and do not architect around a price advantage that has a shelf life.
Do the opposite: build so switching model or vendor is a configuration change. That is the same conclusion the deprecation news points to, arrived at from the cost side.
The part that is not about money
Cheap inference makes it easy to run a model where you do not need one. The cost of a bad automated decision did not fall by eighty percent. Keep the review step.