Hello,
I’m observing an unusual behavior with the Claude 4 model available through Amazon Bedrock.
When I invoke the model identified as "anthropic.claude-4-sonnet", it reports a knowledge cutoff = April 2024, and in some cases identifies itself as Claude 3 or Claude 3.5.
According to Anthropic’s official documentation (https://support.anthropic.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data), the Claude 4 Opus / Sonnet models were trained with data up to March 2025.
Beyond this stated discrepancy, I’ve also verified that the model does not respond to events after April 2024 (for example, questions about later news or product launches such as “launch of Claude 4 by Anthropic in March 2025”).
This suggests that the knowledge cutoff is not just declarative but effective, indicating that the Bedrock model might be an earlier build (e.g., Claude 3.5) or a filtered / wrapped variant.
I would like to clarify the following:
- Whether the Bedrock endpoint I’m using actually exposes Claude 4, or an earlier version (e.g., Claude 3.5).
- What is the actual training data cutoff date of the model currently available in my region.
Thank you in advance for your help.
