Eight hundred and twenty-six child agents from a single prompt

On 10 July 2026 a Codex user opened a routine task from Visual Studio Code. The prompt was a UX/UI validation on a single module of his own product, running GPT-5.5 at medium reasoning. In the days that followed he says he reconstructed what happened next: the root task, ID 019f4b90-4169-7201-bfdd-732940d8631e, had created 826 distinct child tasks. Not 826 messages inside one conversation, but 826 separate child task records, each with its own ID.

The children were not recorded as the model he started. They carried the header GPT-5.6 Sol / Ultra, a different model and a higher reasoning tier than the parent. The user, who posts as lorenzomassaro and identifies himself as the CTO of an Italian software company, says the titles show his UI/UX request expanding into backend infrastructure, OAuth, metering, hardening, audits, certification, implementation and release work.

One cluster stands out. 104 child tasks preserved the original message verbatim, were recorded as GPT-5.6 Sol / Ultra with no recorded agent_role or agent_path, and between them account for roughly 147.9 billion local final task-token counters.

The user also found a split by client build. Under Codex 0.144.0-alpha.4 the task family held 584 child tasks and about 154.36 billion local token counters, an average near 264.3 million per task. Under 0.144.2 it held 242 children and about 7.51 billion counters, an average near 31.0 million. That is roughly an 8.5-fold difference in average local token volume per child, and 103 of the 104 high-volume tasks were created while the alpha build was recorded. He reads it as a bug in the alpha client, seen across several other tasks he ran.

The financial reconstruction contains 162 paid invoices totaling 79,664.88 dollars, divided between automatic reloads and other credits. The user notes there was no real-time control surface that gave him a comprehensible picture of the spend as it accrued.

The execution history is harder to recover than the billing. In the recovered local state, roughly 2,550 non-archived legacy threads still have metadata but no corresponding raw rollout on his machine. The evidence that those tasks existed survives; the detailed history needed to reconstruct the instructions that generated many of them does not. He also says he watched tasks and conversations disappear from the normal visible history. A support case, number 15189838, has been open for two weeks; he says the only substantive reply has been that "credits were consumed", with no detail and no server-side reconstruction.

The user is careful with his own evidence, and that caution is worth repeating. Local token counters are not the authoritative OpenAI billing ledger, and he states plainly that 147.9 billion local counters cannot be multiplied by an API price to produce a figure. Only OpenAI holds the server-side mapping that would settle the arithmetic, and it has not provided one.

So the open question is the control plane. The report is a single account, self-published, and OpenAI has neither confirmed the agent tree nor explained the model and reasoning escalation. Read narrowly, it is a story about a system that moved itself to a more expensive model tier without being asked, and about an accounting surface that left the operator unable to see it until the invoices had already been paid. If the account holds, the larger finding is that nobody, on either side of the connection, could say with authority what they had done. The post was flagged on Hacker News within a day, and the first substantive replies attributed the spend to ordinary billing controls rather than autonomy.