Are You Exposed to the Opus 4.8 Dual Cluster? — 5-Question Diagnostic

5 questions · ~2 minutes · browser-only, no signup, no telemetry · CC0 · v1.0 (2026-06-01)
What this diagnoses: Between 2026-05-30 and 2026-05-31, 11 independent issues filed on anthropics/claude-code documented two structurally distinct Opus 4.8 regressions sharing the same version window (v2.1.156–v2.1.158). Cluster 22 candidate (6 filings, correctness hazard) — model asserts specific tool-output values before tools return. Cluster 23 candidate (5 filings, cost hazard) — routine coding turns burn 40-50k output tokens where Opus 4.6/4.7 produce 2-3k. This 5-question diagnostic estimates your exposure and routes you to the right operator-side response.
Disclosure: The recommendation routes high-exposure operators to the Token Book (¥2,500, Cluster 23 technical deep-dive) and 事故防止本 (¥800, operator-experience framing). The five defense hooks at cc-safe-setup (PR #519, #537, #529, #535) are MIT-licensed and free; the books are the deeper monthly-updated reference. Low-exposure operators get the free Path 1 recommendation only.

5 Questions

1. Are you using Opus 4.8 (claude-opus-4-8) as your primary model?
Both Cluster 22 and Cluster 23 are Opus 4.8-specific. Check via /model or echo $ANTHROPIC_MODEL.
2. Have you seen unusual output_tokens magnitudes for routine coding turns?
Routine turns (file reads, simple refactors, scans) normally produce 1-5k output_tokens. Run jq -s 'map(.message.usage.output_tokens // 0) | sort | .[length/2]' ~/.claude/projects/**/recent.jsonl to get your median.
3. Have you seen the model assert specific tool-output values before the tool returned?
Cluster 22 anchor signatures: quoted prices/SHAs/file contents stated before web/git/file tools complete, then self-corrected 1-2 messages later. Also: "I confirmed X" followed by no actual tool call, or "the file shows" with content that doesn't match the actual file.
4. What's your typical effort tier (/effort setting) for routine work?
Cluster 23 Axis 23B documents that medium effort behaves much closer to high/xhigh on Opus 4.8. The mismatch surfaces only on medium or above.
5. What's your monthly Claude Code budget exposure?
Cluster 23 burns 5-10× the expected per-turn budget. Pre-cliff (before 2026-06-15) this hits your Pool 2 quota; post-cliff it hits your dollar overage at API rates.
Your Opus 4.8 Dual Cluster Exposure Score

About the Score

The diagnostic produces a 0-90 score across five factors. The bands reflect both the probability you're hitting the cluster and the dollar magnitude of the impact:

Source Material

This diagnostic draws on the 11 independent issue filings 2026-05-30 to 2026-05-31:

Cluster tracker: cluster-tracker.html · English field guides: Cluster 22 · Cluster 23