OpenAI confirmed Codex background tasks were silently burning users' usage limits faster than normal
OpenAI (Codex) · Jun 30, 2026 · Coding agent
What happened
Codex users hit their usage limits far faster than normal and OpenAI stood up a war-room response. The company explained that background tasks were consuming excess compute - burning credits even when the user was not actively working - and said a fix was live, resetting caps for affected users to prevent billing errors. Business Insider reported the incident as a vendor-side metering failure in one of the most widely deployed coding agents.
Causal vector
Vendor-side background task execution consuming compute outside the user's visible sessions; metering ran ahead of the user's actual work
Source
Reported by Business Insider. Verified against the primary report.
When the vendor's own metering miscounts, the only control left is yours: an external spend firewall that reconciles billed usage against expected work and alerts on divergence catches silent background burn regardless of whose bug it is.
The six rule types that contain this class of failure
Per-transaction cap
Any single spend above your ceiling is BLOCKED before it moves.
Daily total
Cumulative spend across all agent calls, bounded per day.
Velocity limit
Stops runaway retry loops — the #1 cause of overnight losses.
Merchant allowlist
Only approved destinations can ever receive funds.
Category rules
Flag high-risk classes (crypto, infra, refunds) for review.
Approval threshold
Above a value, the action waits for a human.
Related incidents
Uber burns through its entire 2026 AI coding budget in four months
Uber · Jun 2, 2026
Company hit with a reported $500 million Claude bill after leaving usage limits off
Unnamed enterprise (reported via Axios) · May 28, 2026
Codex extension update left users burning 20% of a weekly token allowance in two hours of ordinary prompts
Codex users (openai/codex issue 14593) · Mar 13, 2026
Don't be the next entry
Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 71 documented failures, one control.