Lemkin says AI resolutions cost 40 to 50 cents each
An investor with a portfolio-wide view says inference costs are not dropping the way founders priced for.
Jason Lemkin says most AI resolutions in B2B cost $0.40 to $0.50 today, and that number is holding even after the model cost savings of the last 30 days or so. He calls it very rough but pretty accurate across 30+ portfolio companies and more. A resolution here is one completed AI task, the unit most support and workflow tools bill against.
If you can charge $2.00+ you are golden
It doesn’t seem to be coming down as much as many of us expected, because quality and competition are driving expectations up equally as quickly
The pricing advice is blunt. If you can charge $2.00 or more per resolution, he says you are golden. If you cannot charge for it, gross margins fall to 60% to 70%, which is well below the 80%-plus software founders and their investors tend to assume.
Why it is not getting cheaper
Lemkin's explanation is not that models stopped improving. It is that the savings keep getting spent. In his words, costs are not coming down as much as many expected because quality and competition are driving expectations up equally as quickly. Every time inference gets cheaper, someone ships a better answer at the same price, and the bar moves.
He also pushes back on the escape hatch most founders reach for, which is that their particular workload is simple. Really simple stuff is cheaper, he says, but it really has to be simple. Really complex stuff costs more, and most things in B2B are not that complex anyway. That leaves a fat middle sitting right around his 40 to 50 cent estimate.
He says the figure tracks with what came up at SaaStr Connect, and that it holds even accounting for recent releases he lists as Jev, Haiku 5.5 and background processing.
The business read
This is one investor's read across his own portfolio, not an audited benchmark, and he labels it rough himself. But it is a wider view than any single founder has, and it cuts against a common plan, which is to launch an AI feature at a thin margin and wait for token prices to bail you out.
The practical version is a pricing question rather than an engineering one. If your product resolves a few hundred tickets a month for a customer on a flat $99 plan, the arithmetic gets uncomfortable fast, and a strong team grinding the cost down is a slower lever than charging per outcome. Lemkin's floor of $2.00 or more per resolution is the number worth testing your own plan against, because it implies roughly a 4x markup on cost rather than the 10x or 20x software people are used to.
The quiet part is the 60% to 70% gross margin case. That is still a business. It is just a different one than the pitch deck said.
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.