[Bug] Subagents billed after monthly spend limit exceeded; false clean review result on agent failures

I'm not sure if this is related or it deserves its own issue, but it appears that there is a bug in model selection and usage.

I have the following case:

• Used available Fable limits per my max plan
• Needed to finish the project I was on and activated usage credits
• I am below my limits still for all models (84%)
• I start a new session and explicitly select opus 4.8
• It seems to have selected Fable regardless and blew $25 before it hit my usage credits limits
• Multiple other sessions have the same behavior - despite choosing Opus, it defaults to Fable and uses credit

Validation ledger (honest)
▗ ▗   ▖ ▖  Claude Code v2.1.198
           Opus 4.8 (1M context) · Claude Max
  ▘▘ ▝▝    ~/code/

 ▎ Fable 5 is back.
 ▎ Until July 7, you can use up to 50% of your plan's weekly usage limit on Fable 5. If you hit your limit, you can continue on Fable 5 with usage credits. Fable 5 draws down usage faster than Opus 4.8. Learn more (https://support.claude.com/en/articles/15424964-claude-fable-5-promotional-access)

❯ /clear                                                                                                                                                                                                                                                                                                                                                                                                                                              

❯ /model                                                                                                                                                                                                                                                                                                                                                                                                                                              
  ⎿  Set model to Opus 4.8 (1M context) and saved as your default for new sessions

You've hit your monthly spend limit. Run /usage-credits to manage your limit and keep using Fable 5 or switch models to continue this chat.

⏺ Switched to Opus 4.8 (1M context) for this session · Fable 5 requires usage credits · /model to change
---
You've reached your Fable 5 limit

You've used your included Fable 5 usage for this week. Continuing on Fable 5 uses usage credits — you have $8.19 in credits.

<img width="1104" height="370" alt="Image" src="https://github.com/user-attachments/assets/303747df-8d54-4ebb-8745-9e0cde1ebb48" />

42 subagents blowing your monthly limit is bad enough. The part that would keep me up is the dead ones returning a "clean" result after they'd been killed. You paid past the cap and got a false pass out of it. That second thing is why I started building a proxy that sits in front of the agents and cuts them off at a budget line instead of trusting the runner to stop itself.

Did the false-clean come from subagents erroring out silently and the parent reading no-findings as a pass? Trying to work out whether that's specific to claude-code or shows up in any fan-out review setup.

The "false clean" result is the more dangerous of these two problems. A code review returning "No findings survived verification" when 6+ of 14 agents silently failed due to billing limits gives exactly the wrong signal - someone merges the PR thinking it passed review.

The spend-limit billing behavior makes some sense as a hard cutoff, but the result rollup should be aware of agent exit reasons. A run where >50% of agents hit the limit should surface as "inconclusive" or "partial", not "clean". Empty findings from errored agents are not the same as empty findings from agents that actually ran.

A simple fix on your end while waiting for Anthropic to address this: check the task/run metadata for agents_error count before trusting a zero-findings result. If agents_error > 0, treat the whole run as inconclusive and block the merge. Not ideal, but it closes the safety gap.

Good detailed bug report - the task IDs with timestamps make this reproducible.

Adding related public pattern context for Anthropic Billing Platform / cost-control triage, because this thread describes subagents continuing to bill after the monthly spend limit was already reached, plus a misleading clean result from failed agents.

Related Anthropic-owned reports in the same cost-control/ledger family:

• Subagent fan-out consumed purchased overage and returned no recoverable deliverable after spend-limit termination: https://github.com/anthropics/claude-code/issues/78231
• Workflow fan-out inherited an expensive tier and triggered automatic usage-credit purchases/card charges without a per-agent or per-workflow spend gate: https://github.com/anthropics/claude-code/issues/68285
• Extra Usage charged/drained while plan capacity still appeared available: https://github.com/anthropics/claude-code/issues/32544

Sanitized public evidence hub collecting related manual-credit, extra-usage, entitlement-routing, auto-topup, and support-routing examples: https://coolak.github.io/anthropic-claude-billing-incident/

Request: please route this failure family to a human Billing Platform / cost-control owner. Agents that cannot run because spend is exhausted should not quietly consume paid buckets, and result rollups should not convert billing-limit failures into clean/empty outcomes without ledger reconciliation.