Option C is correct. For summarization of lengthy documents, exceeding the context window or using large inputs significantly increases token consumption. Text splitting, for example using can break documents into smaller, more manageable chunks. This reduces the number of input tokens per LLM call, directly leading to cost optimization, and is recommended for best search results and LLM response quality. Option A is incorrect because for Cortex AISQL functions, Snowflake recommends using a smaller warehouse (no larger than MEDIUM) as larger warehouses do not increase performance but can result in unnecessary costs associated with keeping the warehouse active. The compute cost for Cortex LLM functions is based on tokens processed, not warehouse size performance. Option B is incorrect because only prevents costs for 'failed' operations by returning NULL instead of an error. It does not optimize the token consumption of 'successful' summarization tasks. Option D is incorrect; Cortex Guard processes additional tokens for its filtering, thus 'increasing' token consumption, not reducing it. Option E is incorrect because setting 'temperature' to 0 makes the output more deterministic, which might improve consistency but does not directly reduce the number of input or output tokens processed for a summarization task.