Breaking

Audit Flags About $1.7M in Mistaken AI Overcharges

June 28, 2026 at 18:31 EDT

AI Agents
Sales & Business
Foundation Models

Vaudit, an auditing startup, reviewed roughly $34 million in AI invoices—mostly tied to Anthropic—and identified about $1.7 million (around 5%) in mistaken overcharges. The bills came from 60 companies between March and June 2026, centered on usage of Anthropic's Claude Code, with some OpenAI usage included as well (techstartups.com).

Mar–Jun 2026 · AI Billing Audit

$1.7M in Errant AI Charges Found Hiding in Customer Invoices

An audit of $34M in Anthropic and OpenAI bills across 60 companies found roughly 5% in mistaken overcharges — driven by agents stuck in "retry storms" and invisible loops. About 80% of disputed charges were credited back.

$1.7M

in mistaken overcharges found

$34M

invoices reviewed · 60 companies

~80%

of disputed charges credited back

How big is the error?

~5% of every dollar billed was charged in error.

$34M

total billed

$1.7M

overcharged (~5%)

The "Retry Storm" Loop

Agent fails task

→

Retries silently

→

Tokens & bill climb

Other causes: invisible background loops, model bait-and-switch, and charges for failed requests — all running before customers notice.

What's working

Real-time, third-party audit tools are already catching and reversing errant charges across OpenAI, Anthropic, Gemini, Bedrock and Vertex AI.

The risk without guardrails

One firm with no spending caps reportedly ran up a ~$500M monthly bill — underscoring the need for limits on agent SDK and tool-call usage.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.