[BUG] Spontaneous language bleed: Russian word in English output

Resolved 💬 2 comments Opened Jun 15, 2026 by jasoncartwright Closed Jun 19, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

During a normal, entirely English-language Claude Code session, the model
emitted a single Russian word — "близко" (Cyrillic for "close/near") — in the
middle of an otherwise all-English sentence, replacing the intended English
word "close". Nothing in the conversation involved Russian or any non-English
text. This is unintended code-switching ("language bleed").

Impact: low severity — cosmetic, not a crash or data/security issue — but it
hurts readability and trust, and would be a real problem if it surfaced in
user-facing generated text.

Environment:

  • Product: Claude Code
  • Model: Claude Opus 4.8 (1M context) — claude-opus-4-8[1m]
  • OS: macOS (Darwin 25.5.0)
  • Claude Code version: 2.1.167 (Claude Code)
  • Client: web

What Should Happen?

The response should stay in the language of the conversation (English). No
foreign-language token should be substituted for an intended English word.
The sentence should have read: "…but you weren't close to hitting it anyway."

Error Messages/Logs

No stack trace or console error was produced — this is a generation-quality
issue, not a crash. The relevant artifact is the verbatim model output:

> One caveat the other way: with the same 1M input context, you didn't gain
> headroom on the input side — but you weren't близко to hitting it anyway.

("близко" = Russian for "close"; it should have been "close".)

Steps to Reproduce

Not reliably reproducible — appears non-deterministic (token-level sampling).
Conditions observed when it occurred:

  1. A long, entirely English session.
  2. Technical/analytical content (a numeric cost-and-token-limit comparison

between two LLM providers).

  1. The slip occurred mid-sentence, on a common high-frequency word.
  2. No Russian or other non-English content anywhere in the prompt or context.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.167 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗