Four hundred support tickets and one question: which field says what. Bulk extraction from tickets, emails, invoices or logs runs at low. OpenAI's guidance puts extraction, routing and classification there, because the task is pattern matching, not judgment. API: reasoning.effort "low". Never send none: Astra returns HTTP 400. Work and Codex: Astra Light. Migrating from none or minimal on an older model? Start here at low and compare. Free template — copy and paste into ChatGPT, Claude, or Copilot.