The 28,000-row test prompt: credit pricing moved the risk from contract to runtime

Thibault Garcia set out to test a new prompt on 10 rows of a large prospecting list in Clay, and it ran on all 28,000 companies in the table ^[1]. That is roughly 2,800 times the intended scope ^[1], and it is the cleanest illustration available of what credit-based pricing has done to software risk: the expensive decision is no longer the signature, it is the click.

Garcia, founder of the go-to-market agency Reachly ^[4], told HubSpot the overrun "cost us a lot of money," adding, "Not too much, thankfully, but it was a lot of money" ^[2]. Clay subsequently introduced a feature to prevent customers from accidentally burning through AI credits ^[3]. Read the sequence carefully. The control existed after the loss, and it protects the next buyer, not the one who paid for the lesson. Anyone evaluating a usage-priced tool today is buying a runtime whose guardrails may still be pending.

The unit economics are not uniform inside a single product, which is what makes scoping errors expensive. In email-finding software, Garcia notes, finding an email might cost one credit, verifying one might cost half a credit, and finding a phone number can run 10 ^[5]. A phone lookup is therefore 20 times the cost of a verification in the same tool ^[2]. Video is steeper: Runway's per-action table prices Google Veo 3.1 with audio at 40 credits per second of video and Nano Banana Pro 4K image generation at 16 credits per image ^[6]. Thirty seconds of that video model is 1,200 credits, the equivalent of 75 4K images ^[3].

Two red flags are worth writing into your evaluation. The first is a stated range with no lookup table. "Some tools are now saying this action could cost you anywhere from one to ten credits, depending on complexity," Garcia says, and without a table translating that into actual cost, he calls it "a very, very big red flag" ^[8]. The second is failure billing. Most tools do not charge a credit when they fail to find what they are looking for, according to Garcia, but not every tool works that way ^[9].

The structural point is about scaling. Per-seat costs track headcount; credit costs track what the AI actually does, which means an entry-tier price tells you almost nothing about what full deployment costs ^[10]. HubSpot's guidance, the third post in its series on credit pricing ^[13], argues that vendors should publish credit rates at multiple commitment levels, because a starting-tier rate will not support a year-two forecast ^[11], and that you should be able to apply published rates to your projected usage and land on the same number the rep gives you ^[12].

Three things to secure before signing, in the contract rather than the roadmap: a per-action credit table, cost previews and confirmation gates on bulk and high-volume actions plus labels on anything that consumes credits ^[7], and a written rule on failed actions ^[9]. Then run the vendor's rates against your own year-two volume ^[12]. If your arithmetic and theirs diverge, the pricing is not the product you were shown.