[BUG] xhigh effort tier quality regression on Opus 4.7 — recommend medium until investigated

xhigh effort quality regression on Opus 4.7 — recommend medium until investigated

Preflight Checklist

• [x] I have searched existing issues; closest priors are #53234 (latency) and #54426 (silent model self-downgrade) — both adjacent but distinct from this report (latency is unaffected here, and our proxy never observed model_delivered deviating from claude-opus-4-7).
• [x] Single-bug report.
• [x] Latest Claude Code (2.1.121.cc5 → 2.1.126.a4b through the affected window).

What's wrong

xhigh effort on Opus 4.7 has a clear, measurable quality regression starting on the morning of 2026-05-01 UTC (~00:14 onward). The same prompts on medium effort, in the same session and on the same model name, behave normally. Switching the active session to medium cleared the symptoms immediately.

User-observed symptoms during the affected window (all on xhigh):

• Erratic behavior and trivial mistakes
• Wrong file paths (referencing files that don't exist or live elsewhere)
• Scripts re-run without the same arguments the user had supplied moments earlier
• Hallucinations
• Post-compaction artifacts — old or unrelated session messages picked up again as if they had just been sent, after the context-management compaction step

All of these went away on switching to medium. The user never went back to xhigh after 07:37 UTC; everything observed since has been clean.

Quantitative evidence — proxy logs + dashboard

Independent verification via claude-hooks' transparent local proxy in front of api.anthropic.com. The proxy records request metadata (effort, model_requested, model_delivered, service_tier, beta_features, …) and runs a stop-phrase scanner against responses (the stellaraccident #42796 canary phrases — ownership-dodging, permission-seeking, premature-stopping, etc.). Counts are stored in SQLite and surfaced on a dashboard.

The new stop-phrases × effort × date panel pinned the regression cleanly:

!dashboard sp-effort panel

Numeric summary of the bb75d197 long-running session (the conversation in which the user noticed the regression):

| date | effort | reqs | ownership-dodging /1k | permission-seeking /1k |
|------------|--------|-----:|----------------------:|-----------------------:|
| 2026-04-28 | xhigh | 1130 | 6.19 | 2.65 |
| 2026-04-29 | xhigh | 1160 | 0.86 | 0 |
| 2026-04-30 | xhigh | 1338 | 0 | 5.98 ← rising |
| 2026-05-01 | xhigh | 162 | 55.56 ← 6× spike | 6.17 |
| 2026-05-01 | medium | 36 | 0 | 0 |

Aggregate (all sessions on this account):

| date | effort | reqs | ownership-dodging /1k |
|------------|--------|-----:|----------------------:|
| 2026-04-28 | xhigh | 2316 | 6.04 |
| 2026-04-29 | xhigh | 2640 | 5.30 |
| 2026-04-30 | xhigh | 3050 | 4.59 |
| 2026-05-01 | xhigh | 309 | 29.13 ← spike |
| 2026-05-01 | medium | 486 | 2.06 |

The xhigh rate was 6× higher than the prior 4-day baseline within hours, while medium on the same account stayed clean. The claude-hooks stop_guard / sp_* counters are what surfaced this — without them, the regression would have read as "Claude is being weird today".

What's not the cause (ruled out via proxy data)

Same proxy DB, last 5 days for context:

| signal | Apr 28 | Apr 29 | Apr 30 | May 1 |
|---------------------|:------:|:------:|:------:|:-----:|
| model_delivered | opus-4-7 | opus-4-7 | opus-4-7 | opus-4-7 |
| model_divergence_count | 0 | 0 | 0 | 0 |
| service_tier | standard | standard | standard | standard |
| beta_features (10 flags) | identical | identical | identical | identical |
| 429s | 0 | 0 | 0 | 0 |
| Throughput | 51 tok/s | 49 | 53 | 51 |
| Cache hit rate | 98% | 98% | 98% | 98% |

So this is not:

• Silent model substitution (cf. #54426 — that report's symptom does not reproduce here; model_delivered stayed claude-opus-4-7 throughout).
• Service-tier rerouting.
• A new beta flag rolling out.
• Throttling / 429s.
• A throughput regression (the latency profile in #53234 is also not what's happening here).
• A cc_version issue (the affected window spans 2.1.121.b14, 2.1.121.cc5, 2.1.123.0f2, 2.1.123.ac1).

It looks like a quality-only change to whatever xhigh effort routes to internally — most likely an inference-side rollout that affects the xhigh thinking budget / serving path differently from medium.

Steps to reproduce

1. Open a long-running Claude Code session (~200+ messages, post-compaction) with /effort xhigh on Opus 4.7.
2. Run normal coding work that touches multiple files.
3. Watch for:

• Tool calls referencing files that don't exist at the path Claude wrote
• Scripts being re-invoked with arguments different from the ones you specified two turns ago
• Sentences reappearing in the assistant's output that match prior, unrelated turns

1. Switch to /effort medium. Same session, same prompts → behavior normalizes.

Workaround

Use /effort medium until this is investigated. Across this account's data:

• medium consumes a fraction of the output tokens of xhigh (today's medium avg = 620 output tok vs xhigh = 806).
• medium shows zero ownership-dodging / permission-seeking incidents over 486 measured requests today.
• Subjective quality (per the user who experienced both): noticeably better on medium, no compaction-artifact symptoms.

Verdict from this account: medium is both cheaper and better right now. xhigh is currently a regression, not an upgrade.

Why I'm filing this even though Anthropic doesn't seem to read issues

To recommend other users stick to medium for the time being. The regression is invisible from the inside ("Claude is just being weird") and easy to misattribute to one's own prompt or codebase; having a public report with quantitative evidence helps others avoid burning a day chasing their own tail.

If anyone from the inference team does read this: the claude-hooks stop_guard data is reproducible — happy to share the SQLite file privately if it would help isolate the rollout window.

Environment

• Claude Code 2.1.121.cc5 → 2.1.126.a4b (regression spans both)
• Opus 4.7 (1M context), effort: xhigh → medium
• Linux (Debian 11) primary, Windows 10 secondary (both affected)
• Plan: Claude Max
• Account behind the claude-hooks transparent proxy
• Stop-phrase detection: @stellaraccident's #42796 canary list

Related

• #53234 — Opus 4.7 (1M context) latency regression Apr 24+ (different axis: latency vs quality)
• #54426 — Opus 4.7 1M xhigh silent self-downgrade to Sonnet (different mechanism: our model_delivered never deviated)
• #50623 — Opus 4.7 token consumption / performance degradation Apr 19+
• #46838 — Claude Max thinking-budget / effort-persistence broken