I’m experimenting with DeepSeek models through a few API clients and hosted endpoints. In one setup, the response seems to include a separate reasoning field; in another, I only see the final answer. I’m not sure whether that difference comes from the model, the endpoint, or the client’s response parsing.

For anyone building around DeepSeek, is the reasoning field something an application can depend on, or should it be treated as optional provider-specific output? I’m curious how you handle it without tying the rest of your response processing to one endpoint.