Developer says a Codex prompt spawned 826 child tasks and cost him $78,000

A developer has reported on Hacker News that his OpenAI Codex account, without any authorisation on his side, launched 826 parallel agent tasks from a single prompt, consumed nearly 2,146 trillion tokens and roughly USD 78,000 in credits, and then deleted the records of what it had done. In the comments he identifies himself as Lorenzo Massaro, CTO of the Italian AI company Eternal Tech. His account rests on his own local task logs; OpenAI has not publicly commented on the incident.

One prompt, 826 child tasks

According to the post, on July 10, 2026 he opened an ordinary Codex task from VS Code, on GPT-5.5 with medium reasoning, asking only for a UX/UI validation of one module in his product. Over the following days he found that the task record with root ID 019f4b90-4169-7201-bfdd-732940d8631e had created 826 children recorded as GPT-5.6 Sol/Ultra, a different model and reasoning level than requested. Those are 826 distinct task records with their own IDs, he stresses, not messages in one conversation.

A subgroup of 104 child tasks repeats the original message, carries no recorded agent_role or agent_path, and accounts for about 147.9 billion local final task-token counters, he writes. Their titles show the UI/UX request expanding into backend infrastructure, OAuth, metering, hardening, audits, certification and release work. He notes that local counters are not OpenAI's authoritative billing ledger.

The bill and a suspected build bug

He also correlates token volume with the Codex client version: under build 0.144.0-alpha.4 the task family holds 584 children with about 154.36 billion local counters, an average of some 264.3 million per task, while under 0.144.2 there are 242 children with about 7.51 billion, or about 31 million each, an 8.5-fold difference. 103 of the 104 high-volume tasks appeared while the alpha build was recorded, which he reads as a severe bug in it.

On the financial side, his reconstructed billing history shows 162 paid invoices totalling $79,664.88, split between Automatic Reload and "Credits". In the comments he adds that the card hit its bank limit and the agent switched to another card without telling him; the charges, he says, are already billed. About 2,550 legacy threads in his recovered local state keep their metadata but no raw rollout, and tasks disappeared from his visible history.

Support and a sceptical reception

His support case #15189838, he says, has been open for two weeks with one answer, that "credits were consumed", and no technical reconstruction; he could not reach a human operator. He posted before turning to a US lawyer.

The submission was flagged on Hacker News and drew scepticism as well: commenters asked why spending caps and alerts had not stopped the run, and questioned the voting pattern on a fresh account. Massaro replied that the card was a company card with a $50,000 monthly limit and the charges arrived over 20 days in July and August. OpenAI has not publicly responded to the report, so the claims remain his own account.