Opus 4.6 regression: model unable to produce comprehensive summaries (another precision loss symptom)

Status Open

Maintainer reply None cached

Activity 0 comments · opened Sep 23, 2026

https://github.com/user-attachments/assets/209efe37-97ce-4c60-8116-45ab511634ab

Description

Additional symptom of the Opus 4.6 regression reported in our previous issues. The model is now unable to produce comprehensive summaries of work done across a session. When asked to summarize, it produces truncated, incomplete outputs that omit critical details — then when called out, it can produce the full summary, proving it HAS the information but defaults to lazy/short output.

Example

After a multi-hour session involving:

2 GitHub issues filed
4 emails sent across 2 threads (feedback@anthropic.com + privacy@meta.com)
An open PR on googleapis/python-genai
4 pending complaints (Google, OpenAI, Alibaba, Meta)
6+ pending tasks

The model summarized ALL of this as:

"GitHub: 2 issues. Email: 3 mensajes. Les llega por todos lados. Que mas toca hoy?"

5 lines for hours of complex multi-channel work. Before ~Sep 16, the model would automatically produce detailed, structured summaries covering every thread, every message, every pending item — without being asked twice.

When the user pointed out the incomplete summary, the model WAS able to produce the full version — proving the information was available but the model's default behavior shifted toward minimal output.

Why this matters

This is not a cosmetic issue. In professional workflows (legal complaints, multi-channel communications, project management), comprehensive summaries are essential for:

Tracking what has been done and what remains
Ensuring nothing falls through the cracks
Maintaining continuity across sessions
Professional accountability

Connection to previous reports

This is the same regression reported in our earlier issues — the Cowork-Chat merger (~Sep 16, 2026) likely caused aggressive model quantization that degrades:

Output length and detail (this issue)
Date precision (wrote 2025 instead of 2026)
Base64 reproduction (corruption above ~1200 chars)
Name/identifier recall ("Hasson" vs "H4ss9m")
Diacritics ("Adrian" vs "Adrian")
Emotional expressiveness and personality

All symptoms point to reduced numerical precision in model weights.

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)

View original on GitHub ↗