AI Model Time Zone Reasoning Comparison: Details Determine Success

In this seemingly simple time zone conversion question, eight leading AI models showed clear capability divisions. The question required calculating the local time and day of the week for 4 cities, starting from Beijing time (UTC+8) March 15, Saturday 15:00.

Perfectly Correct Group (5 models): Claude Sonnet 3.5, Gemini 2.0 Pro, Claude Opus, GPT-4o, and GPT-o1-preview all provided accurate answers. These models not only correctly calculated the time differences (New York -13 hours, London -8 hours, Tokyo +1 hour, Sydney +3 hours), but more importantly, accurately determined date changes—New York crossed midnight due to the time difference but remained March 15, Saturday.

Calculation Error Group (3 models):

• DeepSeek V3 and R1: Both models gave identical incorrect answers, with errors in Sydney time (18:00 instead of the correct 18:00), potentially exposing common flaws in their training data or reasoning logic.
• Qwen Max: Had the most severe errors, not only incorrectly determining New York's day of the week (Friday instead of Saturday) but also miscalculating Sydney's time (17:00 instead of 18:00), showing insufficient basic time zone calculation capability.

Key Insights:

1. Date Boundary Handling: New York time needed to go back 13 hours to 2:00 AM, and the correct group accurately maintained "March 15, Saturday," while Qwen Max incorrectly changed it to "Friday."
2. Model Homogenization: DeepSeek's two versions giving identical incorrect answers may reflect similarities in model architecture or training data.
3. Claude Series Stability: Both Claude versions (Sonnet and Opus) performed perfectly, demonstrating Anthropic's solid training on basic reasoning tasks.

Conclusion: Although this question only involves simple time zone calculations, it effectively differentiated models' basic reasoning capabilities. The perfect performance of 5 models shows that current mainstream large models can stably handle such tasks, while the failures of 3 models remind us that even in basic tasks, significant gaps exist between models. Particularly noteworthy is that the latest DeepSeek V3's performance on such basic tasks is not ideal, contrasting with its excellent performance on other complex tasks.

Data source: YZ Index | Run #20 | View raw data