Codex subagents inherit your main model unless you name one: set a default before big runs

#What changed

Nothing new shipped; this is a usage warning. Builder Austin Lucero (@Austinlucero29) says he started a large Codex workflow on GPT-6.1 Sol at Ultra and asked Codex to use cheap GPT-6 Luna subagents. His local session logs showed 531 subagents on a single thread, nested nine levels deep, and every spawn that didn't name a model ran on Sol Ultra. He says Sol Ultra made up about 80% of the tokens he paid for, and roughly 38,000 credits went in about a day with no warning in the app.

OpenAI's Codex subagents documentation confirms the mechanism behind it: if you don't configure a subagent model or reasoning effort, the subagent inherits the parent agent's model and reasoning effort. Asking for Luna in a prompt only covers the spawns where the model is actually named.

#Who is affected

Codex users who run subagent workflows in the ChatGPT desktop app, the Codex CLI, or the IDE extension, especially with an expensive parent setting such as Sol at Ultra, Max, or xhigh, and anyone whose usage can spill into paid credits after the weekly limit.

#What to do now

• Set agents.default_subagent_model and agents.default_subagent_reasoning_effort under [agents] in ~/.codex/config.toml (for example gpt-6-luna at high), so spawns that don't name a model fall back to the cheaper one instead of your parent's.
• Put model and model_reasoning_effort directly in each custom agent file under ~/.codex/agents/ or .codex/agents/; values in the file take precedence.
• Cap parallel work with agents.max_concurrent_threads_per_session before a large run.
• Watch the subagent list (or /agent in the CLI) and your credit balance during long runs, not after.

#What is not confirmed

The 531-subagent count, the nine-level depth, and the 38,000-credit figure come from one user's logs and are not independently verified. He also says agents.max_depth defaults to 1 but is ignored by the Multi-Agent v2 runtime the desktop app uses; OpenAI's current subagents page doesn't list a depth setting, so whether any nesting cap applies is not confirmed.