Flat pricing sells to the CFO, not the engineer

The Framework

Usage-based pricing looks fair. You charge for what people consume, costs scale with revenue, and nobody overpays. It is the default in AI right now.

It also makes your product impossible to buy inside a company.

Eugene Cheah runs Featherless AI, an inference platform hosting more than 40,000 open source AI models. When the rest of the industry billed per token, he charged a flat monthly rate. The reasoning was not philosophical. Buyers could not answer the one question that unlocks a purchase.

The framework: price so a buyer can defend the number to someone who has not met you.

The Steps

1. Find out who has to approve the spend. Eugene watched non AI native teams stall at exactly this point. His description of the conversation: "they were struggling to tell their boss that said hey, I want to introduce AI to the company and then their CFO will say will be okay, how much will it cost?"

2. Answer that question in one number. The honest usage-based answer was "maybe $10, maybe $1,000, maybe 10,000, I don't know until we try it." Eugene notes the average response to that was "what?" A number nobody can predict is a number nobody can approve.

3. Price the fear, not just the usage. Buyers had read the horror stories. Eugene's version: "I thought I was only going to spend $5 each, charge $500." A fixed rate is protection against bill shock, and that protection is the product feature.

4. Let the constraint pick the model. Eugene was launching with 5,000 models: "I don't want to set up a pricing table for 5,000 models. I'm just going to set a flat rate." Behind the scenes they adjusted for bigger models running slower.

Real Numbers

Subscription price: $10 to $25 a month, positioned for individual chat, requests, and experiments rather than production systems running thousands of parallel requests.

Models at launch: 5,000. A per-model pricing table was never realistic.

Market validation: the major labs followed. Eugene, on where the industry landed: "almost all the major labs now have a fixed pricing plan for coders, for developers. They call this the coding plan. $200 A month." He was early, not wrong.

When It Fails

Flat pricing breaks when you position it for production load. Eugene was explicit that the plan was not for "your entire production system of thousands of requests in parallel." Sell a flat rate into heavy automated usage and your worst customers subsidise nothing.

It also fails when your costs scale linearly with usage and you have no lever to manage that. Featherless could offer it because their hot-swapping meant one GPU serves many models. Without a cost structure you control, a flat rate is a promise you cannot keep.

Note that both models coexist. Eugene's read: "there is a place for both usage based pricing and subscription by a specific capacity based pricing."

Your First Move

Take your pricing page to someone who would have to get budget approval, not to a user. Ask them to tell you what they would put in the request.

If they cannot produce a single number without running your calculator, your pricing is blocking procurement, and no amount of product work fixes that.