HomePrice per Intelligence Point and Cost per Task
Same Job, Who Talks Most: Tokens per Task (Top Effort Setting per Model, CursorBench 3.2, Snapshot Sep 2 2026)
Same Job, Who Talks Most: Tokens per Task (Top Effort Setting per Model, CursorBench 3.2, Snapshot Sep 2 2026)#
The efficiency gain came from talking less, not from a price cut: Fable 5.1 kept its list price, used 30% fewer tokens, and cost 40% less.Draft, pending review
How to read this chart
Each model at its top effort setting. Red is China, blue is the US, brown is Cursor's own model. Cost per task = tokens × list price, so a cheap but talkative model is not necessarily cheap.
Source: Published model pricing and the Artificial Analysis index; FinSight compilation · Updated 2026-09-04
↩ This chart lives on the paged3-intel-price-mon-07