llm-anthropic 0.28 adds default reasoning traces for Claude Fable 5.1 and refusal exception handling

The release of llm-anthropic 0.28 introduces two concrete changes for engineers working with Anthropic’s Claude models. First, reasoning traces, previously optional or hidden, are now visible by default for models that support them, such as Claude Fable 5.1. This shift reduces the need for manual configuration or debugging to inspect model decision-making processes. However, it may also expose internal reasoning that some users might prefer to keep opaque, depending on their use case.

The second change is the addition of a dedicated `llm_anthropic.ClaudeRefusal` exception. This exception specifically handles cases where Claude refuses to comply with a request, such as due to safety or policy constraints. For developers, this provides a more granular way to manage refusals compared to generic error handling. The trade-off is that existing code may need updates to account for this new exception type, particularly if it relies on broader error-catching mechanisms.

The update is narrowly scoped to the llm-anthropic library, meaning it only affects developers who use this specific tooling for Claude model access. There is no indication that these changes alter the underlying behavior of Claude models themselves, only how their outputs and errors are surfaced in the llm-anthropic interface. This makes the update low-risk for most users but also limits its impact to a subset of the Claude developer ecosystem.

While the reasoning traces could improve transparency, they may not be universally useful. For example, applications where latency is critical might see negligible benefit from trace visibility, and some users may find the additional output cluttering. Similarly, the refusal exception, while useful for debugging, does not address the root causes of refusals, only how they are reported. Engineers will need to weigh these trade-offs when deciding whether to adopt the update.