How do you calculate "Credits" from AI usage?

I’ve been talking to a lot of AI founders lately, and one question keeps coming up: "how do you turn raw AI usage into credits your customers actually understand, without losing money on them?"

On paper it sounds simple: 1 credit = X tokens. In practice it gets messy fast:

• Input and output tokens cost different amounts, and the gap changes by model

• One agent run can fan out into several model calls, tool calls, retrieval and voice

• Provider prices change every few months, but your credit price can’t keep changing with them

• Enterprise customers ask for their own conversion rates

• And the question that matters most: are you still making margin on every credit?

Here’s an example of how a simple credit rule can quietly lose money:

Say you sell credits at 1 credit = $0.01 with one rule for every model:

• 1 credit for every 1,000 input tokens

• 1 credit for every 160 output tokens

Request 1: a support question on a standard model
(provider price: $3 per million input tokens, $15 per million output tokens)

• 6,200 input tokens + 900 output tokens

• Credits: 6.2 + 5.6 = 11.8, rounded up to 12 credits ($0.12)

• Provider cost: about $0.03

• Margin: about 73%. Looks great.

Request 2: the exact same question, but the customer switches to a premium reasoning model
(provider price: $15 per million input tokens, $75 per million output tokens)

• Same 6,200 input + 900 output tokens

• Same rule, so still 12 credits ($0.12)

• Provider cost: about $0.16

• You lose about $0.04 on every request

If that customer runs 10,000 requests a month, you lose about $400 every month, while their invoice and your revenue chart look perfectly healthy.

The fix: tie the credit rate to the model’s cost. Make premium-model tokens burn credits 5× faster, and the same request costs 60 credits ($0.60), back to about a 73% margin.

Curious how others handle it:

1. What’s your credit based on: tokens, tasks, outcomes, or something else?

2. Do you price credits from cost plus margin, or from the value delivered?

3. Do you use one conversion rate for every model, or a different rate per model?

4. How do you find out when a customer’s usage has made them unprofitable?

For context, my team has spent months on exactly this problem, and we’re launching @Clocknext here on Product Hunt in the next few days. But I’m genuinely more interested in how you’ve solved it, hacks included.