[Bug] German language regression — English words leak into German responses + grammar errors (2026-01-28)

Resolved 💬 6 comments Opened Jan 28, 2026 by GuidoM21 Closed Mar 6, 2026

Environment

  • Claude Code CLI Version: 2.1.22
  • Model: claude-sonnet-4-5-20250929 (Sonnet 4.5)
  • OS: Darwin 25.2.0 (macOS)
  • Date: 2026-01-28
  • Language setting: German (Deutsch) as primary response language

---

Description

Two distinct failure modes observed in German-language conversations today:

  1. English words leaking into German sentences — no technical justification, pure German equivalents exist
  2. Grammar/declension errors — wrong adjective endings for neuter nouns

Both issues appear to have started today and were not observed in prior sessions.

---

Examples

From another Claude Code session (reported by user):

| Output | Expected | Error Type |
|--------|----------|------------|
| bessere Startlayout | besseres Startlayout | Neuter declension (-e-es) |
| Laut mir kurz in den aktuellen Code | Schauen wir kurz in den aktuellen Code | Completely nonsensical sentence structure |

From this session (self-observed by Claude):

| Output | Expected | Error Type |
|--------|----------|------------|
| komplett garbled | komplett verstümmelt | English word in German sentence |
| kein Reason | kein Grund | English word in German sentence |

---

Impact

  • Undermines trust in German-language outputs
  • Affects professional communication quality
  • Pattern suggests model is under implicit pressure to abbreviate, causing language discipline to degrade

---

Steps to Reproduce

  1. Open Claude Code CLI
  2. Communicate entirely in German
  3. Request code analysis or technical tasks
  4. Observe: English terms and grammar errors appear in responses

---

Expected Behavior

All output in the declared language (German), with correct grammar, declension, and zero unjustified English loanwords.

View original on GitHub ↗

This issue has 6 comments on GitHub. Read the full discussion on GitHub ↗