What's this about?
An OpenAI essay aimed at CFOs argues that software's old success metrics — seats purchased, licenses renewed — don't fit AI. The proposed replacement: "Useful Intelligence per Dollar" — measure work actually accomplished (issues resolved, code shipped, contracts reviewed) against the full cost of producing a successful outcome, not the sticker price per token.
What exactly is wrong with cost-per-token?
A token is the unit AI APIs bill by (a word fragment). Cost-per-token looks objective but hides the denominator that matters: a cheap model may need five attempts and a human review to finish a task a pricier model completes in one pass. OpenAI's framing — cost per *successful task* — is genuinely better accounting, and also, not coincidentally, an argument for buying the expensive frontier model that OpenAI sells.
Why does it matter?
No HN thread on this one — it's a positioning document, but a telling one: the AI sales conversation is shifting from "look at our benchmarks" to "here's how your finance team should measure us," which is what maturing enterprise categories do. Expect "per successful task" pricing arguments (already visible in Kimi K3's $0.94/task marketing) to become the standard battleground.