Short answer: there's no official announcement that Claude's Korean performance has degraded. Still, a growing number of users online say Claude's Korean answers feel less natural than before. This article breaks down where that impression likely comes from and what you can try right now.
Why is this idea spreading?
It's spreading because more people are noticing Claude mix in another language mid-answer or produce sentences that feel awkward in Korean. These reports tend to cluster right after a model transition. When a new model like Claude Opus 5 or Claude Sonnet 5 ships, its phrasing and word choice can differ from the previous model, and some readers interpret that stylistic shift as a quality drop. That said, this is a subjective impression — there's no official material stating that Anthropic has measurably lowered Korean output quality.
Is there an official statement that performance actually dropped?
No. Anthropic's official multilingual support documentation (a developer-facing page explaining how Claude handles different languages) only says that Claude can process most world languages that use standard Unicode characters, that performance varies by language, and that it's particularly strong in widely spoken languages. No figures or statements about a Korean-specific performance drop at any particular point in time appear in that document or any other official channel.
What actually causes Korean answers to feel off
Combining the official documentation with reported cases, the awkwardness people notice generally traces back to four things.
- Tone shifts right after a model switch — a new model can choose different sentence structures or vocabulary than its predecessor. That's a style change, not a performance drop, but it can read as "got worse."
- Language-switching or mixing bugs — Claude may suddenly slip in Japanese or Chinese phrasing mid-conversation in Korean. This is reported especially in conversations mixing CJK languages (CJK — a shorthand for Chinese, Japanese, and Korean).
- Instructions lost as a conversation grows long — as a chat gets longer, Claude summarizes earlier turns to reduce processing load, a step known as context compaction. Early instructions like "always answer in Korean" can get diluted in that process.
- Prompts that don't specify a language — if your prompt (the question or instruction you type) doesn't state your desired language, Claude has to infer it from context alone, and that inference can go wrong.
How to check your language setting
On Claude's web and desktop apps, click the profile icon in the bottom-left corner and choose your preferred language under the "Language" section. This setting affects the interface and default response language, but it isn't an absolute guarantee — the language can still shift depending on conversation content. Voice mode has its own separate setting: go to Settings > General > Voice > Language to set it independently. See the official help article for details.
Fixes you can try right now
If answers feel unnatural, try these three steps in order.
- State the language explicitly — adding something like "please answer only in Korean" directly in your prompt keeps the language more stable than letting Claude infer it from context alone.
- Type Korean in its native script — typing Hangul (안녕하세요) instead of romanization (annyeonghaseyo) tends to produce more natural results.
- Ask for a natural tone specifically — a request like "answer the way a native speaker naturally would, not a literal translation" often noticeably improves the result.
One thing to watch for: in a very long conversation, giving this instruction once may not be enough. Starting a fresh conversation, or restating "please keep answering in Korean" partway through, is the safer approach.
Does this also happen in tools like Claude Code?
Yes. Developers using Claude Code (Claude's terminal-based tool for writing and editing code) have also reported the language suddenly switching mid-conversation. Some report that adding a rule like "always respond in Korean" to a project instructions file (such as a CLAUDE.md file placed in the project folder) helps the language instruction hold up better as the conversation grows longer. Keep in mind this isn't an officially guaranteed behavior — it's a workaround users have found effective in practice.
Frequently Asked Questions
Q. Is there an official statement that Claude's Korean performance actually dropped?
No. Anthropic's multilingual support documentation only states that performance can vary by language in general — no material states that Korean performance dropped at a specific point in time.
Q. Is language mixing mid-conversation the same issue?
The cause may differ. Language mixing appears more closely tied to language-switching bugs, or to context compaction diluting instructions in long conversations, than to an actual drop in Korean quality.
Q. Does changing the language setting fix it immediately?
Not with full certainty. The profile language setting only sets a default, so it's safer to also state your desired language and tone directly in your prompts.
Q. Does Korean tone change when Claude switches to a new model?
It can. A model switch can change sentence structure and word choice — that's a stylistic difference specific to the model, not a performance drop.
If Claude's Korean suddenly feels worse, checking how you specify the language, how long the conversation has run, and whether a model switch just happened will get you closer to a real fix than assuming performance itself declined.