Opus 4.7 regression: weaker keyboard-language typo detection + less precise memory/MD file writes vs Opus 4.6

Environment

• Claude Code version: 2.1.118
• Model: claude-opus-4-7 (1M context)
• Platform: Windows 11 Pro (win32)
• Shell: bash

Summary

After using both Opus 4.6 and Opus 4.7 extensively on the same workflows, I consistently observe two regressions in 4.7:

1. Weaker keyboard-language typo auto-detection

When I accidentally type Arabic characters while my keyboard is in Arabic mode but I meant English (or vice versa), Opus 4.6 recognizes the mistake and replies in the intended language. Opus 4.7 does not — it responds as if the garbled input were intentional, or asks for clarification instead of inferring.

This matters for bilingual Arabic/English users (very common in the GCC). 4.6's behavior felt "smart"; 4.7 feels literal.

2. Less precise memory / markdown file writes

On the same project with the same CLAUDE.md hierarchy and auto-memory system, Opus 4.6 maintained memory files cleanly and updated MEMORY.md indices correctly. Opus 4.7 more frequently:

• forgets to add a pointer line in MEMORY.md after writing a new memory file
• saves content that duplicates existing memories instead of updating
• writes longer, less structured entries (vs the one-line - [Title](file.md) — hook convention)

Expected

Opus 4.7 should match or exceed 4.6 on instruction-following for structured file operations (memory/MD saves) and on implicit bilingual-context inference.

Actual

4.7 regressed on both. Matches documented community reports of weaker zero-shot instruction adherence and long-context recall in 4.7 vs 4.6 (see The Register, MindStudio, Medium reviews from Apr 2026).

Workaround

Switching to Opus 4.6 or Sonnet 4.6 for memory-heavy and bilingual sessions.

Impact

Bilingual users and users with disciplined memory/MD conventions are getting worse results on the flagship model than on its predecessor.