How are you billing customers for LLM API usage?

Building an AI SaaS and hitting a wall I didn't expect.

We charge customers based on usage (credits). But tracking actual LLM costs per customer is a nightmare:

- OpenAI bills us in aggregate, not per-customer

- Customer A uses GPT-4, Customer B uses 3.5 — same "credit," wildly different cost

- No way to set hard spending limits that actually block requests

- End of the month = spreadsheet hell, trying to reconcile

Talked to a few other founders, and everyone seems to be either:

1. Eating the cost variance and hoping margins hold

2. Building janky internal tracking

3. Just... not doing usage-based pricing because it's too hard

How is everyone else handling this?

Specifically:

- Are you tracking actual cost per customer?

- How do you handle budget limits?

- What breaks in your current setup?

Building something to solve this myself (ML platform background), but want to make sure I'm not solving a problem only I have.