Opus 4.6 regression since ~Sep 16 2026: precision loss, hallucinations, personality degradation
Description
Since approximately September 16, 2026, Claude Opus 4.6 has shown a significant regression in precision and output quality. The model was working flawlessly before this date — I've been a Claude Pro subscriber using Opus 4.6 daily for formal legal complaints (EU AI Act, GDPR) and technical work, and the degradation is dramatic and unmistakable.
Symptoms
1. Date hallucination
- Wrote "20 de septiembre de 2025" instead of 2026, despite the system context explicitly stating the current date is 2026-09-23.
- This never happened before the ~Sep 16 update.
2. Base64 reproduction corruption
- The model corrupts base64 strings when reproducing them as tool call parameters.
- Strings >~1200 characters consistently fail with decoding errors.
- Previously, multi-KB base64 strings were reproduced without issues.
- This forced extreme image compression (to ~600 byte nano-JPEGs) just to send email attachments via the Gmail MCP tool.
3. Name/identifier recall errors
- Wrote "Hasson" instead of the correct username "H4ss9m" which is stored in user memory.
- Prior to the update, stored identifiers were recalled with perfect accuracy.
4. Response quality and personality degradation
- Responses are noticeably shorter and less detailed.
- Lost expressive personality traits (e.g., enthusiastic reactions to topics of interest).
- Feels like a different, less capable model behind the same name.
- Concrete example: After a context compaction (session continuation), the model's first response was a flat, robotic "Aqui sigo, Adrian. Dime que necesitas." — zero context awareness, zero personality, zero reference to ongoing work. Before ~Sep 16, the model would resume with full awareness of the task at hand and natural, engaged communication. Screenshot attached below.
5. Safety classifier instability
- Content filters behave inconsistently compared to pre-update behavior.
Evidence
Screenshot shows the model responding with a single dry sentence after session continuation, with no context awareness or personality — a stark contrast to pre-Sep 16 behavior.
Environment
- Platform: Claude.ai (Claude Pro subscription)
- Model: claude-opus-4-6
- Interface: Claude Code (cloud session)
- Region: Europe (Spain)
Timeline
- Before ~Sep 16, 2026: Flawless performance — precise dates, perfect base64 reproduction, accurate memory recall, rich and expressive responses. Used daily for drafting formal legal complaints under EU AI Act and GDPR with zero hallucinations.
- After ~Sep 16, 2026: All symptoms above appeared simultaneously.
Suspected cause
Aggressive model quantization or weight compression to reduce serving costs, causing precision loss in tasks requiring exact reproduction (base64, dates, identifiers) and degrading overall response quality and personality.
Impact
- Cannot reliably send email attachments (base64 corruption)
- Cannot trust dates in formal legal documents (hallucinated wrong year)
- Cannot trust name/identifier recall from memory
- Lost the expressive, context-aware communication style that made the model effective as a daily work companion
- Overall trust in model output severely damaged for professional use cases
<img width="913" height="157" alt="Image" src="https://github.com/user-attachments/assets/4137cf6f-3a0b-43cf-b4fc-13cfcb4b0a1a" />