Detailed Analysis
A Reddit post in Italian raises a user-reported quirk with Claude Code running on the Sonnet model: the tool's internal reasoning output appears to switch languages mid-stream, shifting from Italian to English to Spanish without any apparent user instruction to do so. The poster asks whether this is a known issue, why it happens, and whether it should be considered a problem or simply a harmless side effect of how the model operates. No official Anthropic documentation or research context accompanies the post, meaning this is an anecdotal, community-sourced observation rather than a confirmed or explained bug.
This behavior is plausible given how large language models like Claude generate extended reasoning traces (often called "thinking" or chain-of-thought tokens) before producing a final answer. These intermediate tokens are typically less constrained than the polished final output, since the model treats them as a scratchpad for working through a problem rather than as user-facing prose. Because Claude is trained on massive multilingual corpora and internally represents concepts in a shared latent space that isn't strictly tied to any one language, it can default to whichever language most efficiently encodes a given thought at a given moment—English is heavily represented in training data and often serves as a kind of computational "pivot" language, while Spanish or Italian tokens might surface when the model is processing input terms, proper nouns, or concepts that map more naturally to those languages. This is consistent with published interpretability research from Anthropic itself, which has shown that Claude models often reason in an abstract, language-agnostic conceptual space internally, with the choice of surface language being somewhat incidental to the underlying computation.
Whether this is "good" or "bad" depends on what the user values. For raw task performance, code correctness, and factual accuracy, mixed-language reasoning traces are unlikely to matter much, since the final answer is usually normalized to the user's preferred language and the reasoning process is an internal step rather than the deliverable. However, for users who rely on reading the chain-of-thought to audit, debug, or trust the model's reasoning—especially non-English speakers who want transparency in their native language—inconsistent language switching can be disorienting or reduce perceived reliability, even if it has no functional impact on output quality. It also raises minor concerns about token efficiency and latency, since switching representational modes mid-reasoning could, in theory, introduce small inefficiencies, though this is speculative without benchmarking data.
More broadly, this anecdote reflects a recurring theme in the deployment of frontier multilingual AI models: the gap between how these systems compute internally and how they present themselves to users. As reasoning-capable models like Claude are increasingly used by non-English-speaking developers worldwide, questions about language consistency, localization quality, and the transparency of extended thinking traces are likely to grow in importance. Anthropic and other labs will likely need to decide whether to enforce stricter language consistency in visible reasoning outputs—potentially at some cost to raw model flexibility—or to treat this as an acceptable artifact of multilingual training that doesn't meaningfully degrade the user experience for the vast majority of use cases. For now, the phenomenon described in this post appears to be a byproduct of Claude's underlying architecture and training rather than a bug in the traditional sense, though it underscores ongoing challenges in making advanced AI reasoning both effective and legible across languages.
Read original article →