Opus 4.6 regression: catastrophic output quality loss after context window summarization

Bug description

When a long Opus 4.6 session exceeds the context window and prior turns are summarized, the model's output quality drops catastrophically. This is distinct from the gradual precision loss reported in #96182 — this is an abrupt, severe degradation that occurs at the moment of summarization.

Symptoms after context summarization

1. Model produces empty or minimal responses where it previously gave detailed, contextual answers
2. Complete loss of contextual awareness — the model cannot reference prior work, decisions, or established patterns from the session
3. Summaries of prior work are truncated — hours of collaborative work reduced to 3-5 flat sentences (documented in #96289)
4. Responses become generic and formulaic — losing the specificity and depth that characterized pre-summarization outputs

Steps to reproduce

1. Start a long Claude Code or Claude Pro session on Opus 4.6
2. Engage in sustained, multi-step work (research, issue filing, code review)
3. Continue until context window summarization triggers
4. Observe: the model's next response shows severe quality degradation
5. Subsequent responses remain degraded for the rest of the session

Expected behavior

• Context summarization should preserve enough information for the model to maintain output quality
• The model should gracefully degrade, not catastrophically fail
• Key context (names, project details, prior decisions, established tone) should survive summarization

Actual behavior

• Output quality drops to near-zero immediately after summarization
• The model cannot maintain established patterns or reference prior work
• Critical details are lost (e.g., proper nouns, accented characters, project-specific terminology)
• The session effectively becomes a new conversation with no history

Impact

This makes Opus 4.6 unreliable for any sustained work session. Claude Code users who rely on long sessions for complex tasks (multi-file refactoring, research, iterative development) cannot trust that their work context will survive summarization.

Timeline

This behavior began around September 16, 2026. Prior to that date, context summarization preserved significantly more information and output quality was maintained post-summarization.

Related issues

• #96182 — Opus 4.6 precision loss, hallucinations, personality degradation
• #96283 — Root cause: infrastructure changes (~Sep 16) broke model precision
• #96289 — Model unable to produce comprehensive summaries
• #96292 — Consumer rights complaint (Directive 2019/770)

Reporter

• Adrián García Casado
• Claude Pro subscriber
• GitHub: adri22235