[Bug] Session/weekly token limits deplete disproportionately fast with Opus orchestrator + Haiku subagents
Bug Description
Subject: Max 20x — session + weekly limits drain disproportionately fast when Opus is used as pure orchestrator with Haiku subagents
Plan: Max 20x ($200/month)
Client: Claude Code CLI (latest)
Setup:
- Main model: Opus 4.7 — used STRICTLY as an orchestrator. It does not write code, does not read large files, does not do heavy implementation work.
- All custom subagents: explicitly pinned to
model: haikuvia frontmatter. They do 100% of the actual work (file reads, code edits, searches, tool calls). - Opus's job in my workflow is limited to: receive the user request, delegate to the right Haiku subagent, review the returned summary, delegate the next step.
Issue:
After ~40 minutes of this workflow, the 5-hour session window shows 22% consumption. The weekly limit depletes at a rate that would exhaust it in 2-3 days of normal work, not 7. This is consistent with the behavior reported in GitHub issues #8449 and #26271, which remain unresolved.
Why this is wrong:
With Opus acting purely as a router — short delegation prompts in, short summaries out — the Opus token consumption per turn should be minimal. The bulk of tokens (file contents, grep results, code generation) lives in the Haiku subagents' isolated context windows and should be billed at Haiku rates. If the numbers are correct, then either:
(a) Opus orchestration is being billed as if it were doing the full work of the subagents, or
(b) Haiku subagent invocations are silently being routed to a more expensive model, or
(c) There is a bug in how session/weekly consumption is calculated, as already reported in #8449 and #26271.
There is no way for me to tell which, because the usage dashboard shows a single aggregate number with no per-model breakdown.
Environment Info
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.118
- Feedback ID: 326a7987-ee3d-4cee-8822-3205ffefa92c
Errors
[{"error":"Error: NON-FATAL: Lock acquisition failed for /Users/lohonyaigyozo/.local/share/claude/versions/2.1.118 (expected in multi-process scenarios)\n at ai_ (/$bunfs/root/src/entrypoints/cli.js:2756:2177)\n at IK8 (/$bunfs/root/src/entrypoints/cli.js:2756:1257)\n at processTicksAndRejections (native:7:39)","timestamp":"2026-04-23T18:17:38.431Z"}]This issue has 10 comments on GitHub. Read the full discussion on GitHub ↗