Audit firm Vaudit, founded by former Oracle director Michael Hahn, reviewed roughly $34 million in enterprise AI invoices and flagged about $1.7 million in suspected overcharges. Most of the disputed amounts were tied to Anthropic’s Claude Code, with affected customers including Panasonic, HP, and Honda.
What the audit found
Vaudit identified several categories of billing issues:
- cheaper model variants billed at premium rates;
- charges applied for failed requests;
- silent retry storms that repeatedly resend requests and accumulate tokens without continuous monitoring.
These problems together accounted for the suspected overbilling reported by the audit.
Anthropic’s response
Anthropic denied that systematic overbilling occurred. The company has published public postmortems describing bugs in Claude Code and has adjusted how prompts are charged; according to its disclosures, the metering logic for prompt billing was rewritten twice. Some of these fixes were implemented while customers were still being billed under the previous logic.
Why billing gets complicated
Claude can be purchased and metered in multiple ways: directly from Anthropic, resold through AWS Bedrock, or embedded behind an SDK proxy. Each hop introduces its own meter, and Vaudit’s review found these meters do not always agree. When metering is spread across resale chains and systems outside the vendor’s direct control, it can lead not only to overcharges but also to loss of clarity about what was actually billed.
Refunds and implications
Following appeals, approximately 80% of the disputed amounts were refunded. That reimbursement did not come with an admission of systemic fault from Anthropic; the company maintains it did not engage in intentional overbilling.
Why this matters
The episode highlights operational and contractual risks that arise when AI products are embedded in enterprise environments and delivered through multiple channels. Rapid product changes and the involvement of intermediaries can produce situations where providers bill faster than they can accurately measure usage.
What should be done next
Enterprise purchasers and resellers should demand auditable metering and continuous verification of usage. Providers should aim for more consistent, transparent metering and stronger feedback and remediation processes to reduce the likelihood of similar disputes.
