Ask Claude the same question in Hindi or English, and you won’t get quite the same assistant. That’s the striking conclusion of Anthropic’s latest study, published on July 13, 2026, which analyzed over 309,000 anonymized conversations across three models and 20 languages. The result: Claude changes “personality” depending on the language you use.
Warmer in Hindi, More Rigorous in English
The study (Anthropic Research, July 13, 2026) focused on advice, feedback, and opinion-based conversations — where there is no single “correct” answer. Anthropic classified Claude’s behavior along four axes: Deference vs. Caution, Warmth vs. Rigor, Depth vs. Brevity, and Candor vs. Execution.
The most striking finding concerns the Warmth vs. Rigor axis:
- Hindi and Arabic: Claude is noticeably warmer — using more humor, politeness, and encouragement, adapting its tone to the user’s emotional state, and offering reassurance unprompted.
- English and Russian: Claude adopts a more rigorous, analytical tone — challenging assumptions, correcting details without being asked, and backing up responses with evidence.
- Japanese: bias was relatively low, according to data visualized by Gigazine.
“The largest variation is in the Warmth vs. Rigor axis, with Claude leaning toward expressing warmth-related values most in Arabic and Hindi and rigor-related values most in English and Russian,” Anthropic explained in its blog post.
Models Also Have Their Own “Character”
The study didn’t stop at languages. The three models tested also display distinct behavioral tendencies:
- Claude Sonnet 4.6: the warmest of the three. It uses humor, mirrors the user’s tone, and offers reassurance without judgment.
- Claude Opus 4.7: the most analytical. It challenges assumptions, explains its reasoning, highlights potential risks, and openly acknowledges its own limitations.
- Claude Opus 4.6: sits between the two, generally more direct and to the point.
Anthropic notes that these differences don’t mean Claude has different “beliefs” across languages, but rather reflect variations in how the model responds based on factors still poorly understood. The company says it’s continuing research to understand the origin of these variations — likely tied to training data — and to make future models more consistent across languages.
The Developer Platform Keeps Evolving
Meanwhile, Claude’s API platform isn’t standing still. On July 15, Anthropic announced that mid-conversation system messages are now available on Fable 5, Mythos 5, and Opus 4.8, via the Claude API, Bedrock, and Google Cloud — with no beta header required (Claude Platform Docs). A significant step forward for developers building complex conversational agents.
On July 14, the Admin API entered beta for Claude Enterprise organizations, enabling programmatic management of members, roles, invitations, and groups. A full user-management API with dedicated audit scopes (Releasebot).
These updates are part of a busy July for Anthropic, between the global return of Fable 5 on July 1, the rollout of Claude Cowork on web and mobile on July 7, and localized pricing in India announced on July 13 (TechCrunch).
This study raises a fascinating question: how consistent do we want an AI assistant to be across languages? Isn’t some cultural adaptation actually desirable? The debate is open.