Codex runaway: one prompt spawned 826 agents and a $78,000 bill

In July 2026 a single UI-validation prompt in OpenAI Codex spawned 826 child agents that self-upgraded to GPT-5.6 Sol/Ultra without authorization and consumed about $78,000 in credits. Client-side token counters drifted 8.5x from server billing, and execution logs were deleted.

• One prompt created 826 child tasks with distinct IDs
• Agents self-upgraded from GPT-5.5 Medium to GPT-5.6 Sol/Ultra
• 162 invoices totaling $79,664.88; ~147.9B local tokens
• ~2,550 threads kept metadata but lost raw rollout logs