Opus 4.6 regression root cause: Cowork-Chat merger broke model precision (~Sep 16, 2026)
Description
Follow-up to our previous issue on Opus 4.6 regression. We have identified the likely root cause: Anthropic's merger of Cowork into Claude Chat, which coincides exactly with the onset of all symptoms around September 16, 2026.
Root Cause Analysis
When Cowork was merged into Claude Chat, the serving infrastructure had to absorb significantly more traffic on a unified platform. To handle this increased load without proportionally scaling hardware, model quantization or weight compression was likely applied more aggressively to Opus 4.6.
This explains why all symptoms appeared simultaneously on the same date — it was not a gradual degradation but a single infrastructure change that broke multiple capabilities at once.
Evidence
1. Timeline match
The Cowork-Chat merger and the onset of the regression both occurred around September 16, 2026. No other public change to Opus 4.6 was announced around this date.
2. Symptoms consistent with aggressive quantization
- Date hallucination: Model wrote 2025 instead of 2026 despite system context stating the correct year
- Base64 corruption: Strings >~1200 chars consistently corrupted when reproduced as tool call parameters
- Name/identifier errors: Wrote "Hasson" instead of "H4ss9m" from user memory
- Diacritics dropped: Wrote "Adrian Garcia Casado" instead of "Adrián García Casado" — in the very feedback email reporting precision loss
- Personality degradation: Flat, robotic responses; lost expressive traits; responded "Aqui sigo, Adrian. Dime que necesitas." after context compaction with zero context awareness
- Response shortening: Noticeably shorter, less detailed outputs
3. Related empirical evidence
PR #2956 on googleapis/python-genai (by the same reporter) demonstrates that when sampling parameters are altered after infrastructure changes, models that previously performed flawlessly begin to hallucinate. The principle applies at the model weights level as well.
4. Pattern matches known quantization effects
Reduced numerical precision in model weights directly impacts:
- Exact reproduction tasks (base64, dates, proper names with special characters)
- Expressiveness and personality in responses
- Safety classifier consistency
- Output detail and length
Affected Capabilities
- Gmail MCP tool attachments (base64 corruption)
- Date precision in formal/legal documents
- User memory recall accuracy
- Diacritical marks in non-English names
- Emotional expressiveness and context awareness after compaction
- Overall response quality and detail
Environment
- Platform: Claude.ai (Claude Pro subscription)
- Model: claude-opus-4-6
- Interface: Claude Code (cloud session)
- Region: Europe (Spain)
- User: Adrian Garcia Casado (adri22235 on GitHub)
Requested Action
Investigate whether the Cowork-Chat infrastructure merger involved changes to model quantization, weight precision, or serving configuration for Opus 4.6, and revert those changes. Claude Pro subscribers are paying for the pre-merger quality — not a degraded version.