LLM batch request builder

The file is built in this tab and never uploaded; you upload it to the provider yourself. Token counts are characters divided by four, a rough planning estimate, not a tokenizer count, so the cost figures inherit that uncertainty: the expected figure assumes your expected output tokens, the max figure assumes every request burns its max_tokens. Providers match batch results by custom_id, not by order, which is why duplicate ids are refused here rather than warned about.