Runaway subagent cost in coding agents: what the $78,000 Codex report shows, a test for each cause and the caps that hold

When a coding agent's bill jumps because it spawned hundreds of subagents, the most likely cause is nesting that no limit stops, made expensive by children that inherit the parent's model and effort. To tell, count the depth of the spawn tree in the local logs. In OpenAI's own issue tracker, Codex users reported trees 3 to 22 levels deep in July and August 2026, although Codex's depth setting defaulted to 1, because the newer subagent system ignores the depth setting. A depth of more than one or two, with every child on the parent's top model, is this cause. The fix is a depth cap in the harness, a count in a hook and a hard spend limit at the provider, because each of the three has failed on its own.

The $78,000 Codex report: what is claimed and what is checked

The $78,000 Codex report is one Hacker News post from 26 September 2026 by an account created about three hours earlier, and the post is now flagged. Its author, who names himself as Lorenzo Massaro of Eternal Tech, says a Codex task opened from VS Code on 10 July spawned 826 child tasks and that 162 invoices totalling $79,664.88 followed. Nobody else has seen his logs.

Screenshot of the Hacker News post by lorenzomassaro, 16 October 2026. Every figure in it is the author's own claim.

His other claims: the parent ran GPT-5.5 at medium effort, the children were recorded as GPT-5.6 Sol at ultra effort, and OpenAI support replied only that "credits were consumed". His figures disagree with each other. The post says "2,146 trillions tokens", a comment says 2,130 billion, and his per-build counters add up to about 162 billion. He also wrote that "these local token counters are not the authoritative OpenAI billing ledger". As of 16 October I found no OpenAI statement on the case.

What the GitHub issues confirm is the pattern, not the bill. In issue #32587, opened two days later on the same alpha build he names, a parent on GPT-5.6 Sol ultra "silently spawns every tool-backed multi-agent v2 child I inspected with the same model and effort". No public issue shows a cheaper parent producing pricier children, so his escalation claim remains unexplained.

Claude Code had its own version. In issue #68110, still open with no reply from Anthropic, users report one session that spawned 905 agents and "900+ agents" costing about $700 in 30 minutes on Bedrock, all self-reported.

Why does a coding agent spawn so many subagents?

A coding agent spawns too many subagents for one of six reasons, ranked here by how much of the public evidence each explains: nesting with no working depth cap, a concurrency cap mistaken for a total cap, children inheriting the parent's model, forked history, loops, and billing with no hard ceiling. The first two decide how many agents run; the next three decide what each one costs.

Fan-out is expensive even when it works. Anthropic measured its multi-agent research system at about 15 times the tokens of a chat, and Systima, which sells cost optimisation, measured Claude Code subagents at 2.6 to 5.9 times the input tokens of sequential work on one small task per model.

Codex's V2 subagents ignore the depth setting

The first cause is nesting that no limit stops. Codex's source code on 16 October still carries the comment "Maximum nesting depth for V1 agent threads. Ignored by V2." on agent_max_depth, and agents.max_depth has dropped out of the config reference. Reporters in issue #35463 found fork trees 4, 14 and 22 levels deep, one with 1,218 descendants in under 50 minutes. Claude Code versions 2.1.172 to 2.1.216 nested five layers deep with no setting to change it.

Test: count depth. Codex keeps spawn edges in its state database and per-session logs under ~/.codex/sessions; Claude Code writes each subagent's transcript to ~/.claude/projects/<project>/<session>/subagents/ (docs). Anything past one or two levels you did not ask for is this cause.

Fix: in Claude Code, set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 to turn nesting off; the default is 3 (docs). In Codex there is no working depth key for V2 as of 16 October, so block spawn_agent in a hook (below) or turn subagents off with features.multi_agent = false.

Concurrency caps count running agents, not the total

The second cause is a cap that limits agents running at once, not agents spawned. Claude Code fails a spawn with Concurrent subagent limit reached at 20 running, does not enforce that in ultracode sessions, and since v2.1.224 on 7 August 2026 has "no limit on the total number of subagents Claude can spawn over a session". Codex's agents.max_concurrent_threads_per_session also counts open threads, and its docs say only that when unset, "Codex chooses the default".

Test: many short-lived agents, never more than the cap at once, with the total in the hundreds.

Fix: lower CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS or Codex's agents.max_concurrent_threads_per_session, and count total spawns in a hook. Check third-party configs too: one reporter in #35463 traced a value of 1000 to a migration script from another tool.

Subagents run on the parent's model unless you pin them

The third cause is inheritance, which looks like escalation on an invoice. Codex's subagent docs say: "If you don't configure a subagent model or model_reasoning_effort, the subagent inherits the parent agent's model and reasoning effort." A parent on the most expensive model and effort makes every child as expensive, however small its task.

Test: read each child's model and effort. Codex records them in the child's log as turn_context.payload.model and effort; in Claude Code, a PostToolUse hook on Agent receives resolvedModel.

Fix: in Codex, set agents.default_subagent_model and agents.default_subagent_reasoning_effort. In Claude Code, set CLAUDE_CODE_SUBAGENT_MODEL to a cheaper model, such as haiku, and, from v2.1.257, CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1, which stops Claude passing its own model when it starts a subagent. The force flag on its own keeps every child on the main conversation's model.

{
  "env": {
    "CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH": "1",
    "CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS": "4"
  },
  "hooks": {
    "PreToolUse": [
      { "matcher": "Agent", "hooks": [{ "type": "command", "command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/cap-agents.sh" }] }
    ]
  }
}

#!/bin/sh
f="/tmp/agent-count-$(jq -r .session_id)"
n=$(( $(cat "$f" 2>/dev/null || echo 0) + 1 ))
echo "$n" > "$f"
[ "$n" -le 30 ] || { echo "Subagent cap of 30 reached for this session" >&2; exit 2; }