I am writing this because I spent the equivalent of $50 in API compute to discover that Claude had already billed me — through errors of its own making — for roughly $300 I should not have paid.

I want you to read that sentence again.

Your AI made mistakes. Your platform billed me for them. I had to commission a forensic session just to find out it was happening.

What I Found

I have been a Max subscriber since 10 April. I run Cowork sessions daily — building an AI marketing consultancy, working with your product in good faith, leaning in hard.

Over 59 days I accumulated 1,601 Cowork sessions. I ran a script this week against every audit.jsonl file your platform stored on my machine. The total API-equivalent spend: $38,633.

I am not naive. I know Max is a flat subscription. I know I am not being handed an invoice for $38,633. But I am not stupid either. That number represents real compute consumed. Real capacity used. Real rate limits hit. And a material chunk of it was consumed not by my work — but by Claude's failures billing themselves back to me.

Exhibit A: You Billed Me Three Times for Your Own Rate Limit

Here is a verbatim extract from your own audit log. 8 June 2026, 13:33 UTC:

One user message. Three billable assistant turns. Then — at line four — your own platform logging a rate_limit_event.

The sequence is not ambiguous. Claude hit your rate ceiling. Your platform retried. Twice. Each retry re-wrote 82,000 tokens of context cache at full input price. You logged the rate limit. You billed for all three turns anyway.

That is $0.77 for a single four-word message. The error is yours. The bill is mine.

This pattern recurs. At 13:35:02, the same thing:

$0.44. One message. Your rate limit. My cost.

Exhibit B: $6.15 to Open an Old Chat

On the same morning, I opened a conversation I had used two hours earlier. I typed two words: "save to memory."

Your platform re-wrote 407,069 tokens of prompt cache at Opus rates to reconstitute context it had already processed.

$6.15. To reopen a chat.

There is no warning before this happens. There is no UI indicator that says: this conversation has gone cold; resuming will cost the equivalent of 200 fresh API calls. The charge is silent and automatic. The user has no agency.

Exhibit C: Parallel Tool Calls Billed as Independent Turns

When I ask Claude to search for two files simultaneously — a standard efficiency — each parallel tool call generates a separate audit entry, each with a full context cache write:

Three turns. One logical action. This is not a user error. This is how your architecture bills for its own parallelism.

The Meta-Cost

Here is the part that should concern you most.

I did not know any of this was happening. Your product has no cost visibility. No session-level spend estimate. No warning when a cold Opus cache-write is about to cost six dollars. No alert when your own rate limiter triggers a retry loop at my expense.

What I Am Not Saying

I am not saying your product is bad. I use it daily. I am building a business on it.

I am not saying the $38,633 is money you owe me. I understand Max pricing.

I am not even saying Claude's mistakes are malicious. Rate limit retries are an engineering decision. Cache warming is a design choice. Parallel tool billing is an architecture pattern.

What I Am Saying

You are billing users for platform failures without telling them.

The rate_limit_event is in your own log. You know when you hit the ceiling. You retry anyway. You bill for the retries. Then you log the event that proves what happened.

That is not a grey area.

A user on your most expensive consumer tier, using your product in exactly the way you intend, is being charged for errors generated by your infrastructure. The errors are logged. The billing happens anyway. The user has no visibility until they build their own forensic tooling.

The product is powerful. The trust relationship this creates is not.

What I Would Like

I found all of this. I documented all of this. I am sharing all of this publicly because I suspect I am not the only one.

The logs are on my machine. The receipts are real. The question — as I put it to you directly — is this:

Does it seem fair to you that I had to spend $50 to discover you had been charging me for $300 worth of your own mistakes?

I would genuinely like an answer.