Enable extended thinking when the task requires careful multi-step reasoning, not just quick synthesis. Typical triggers are: debugging complex issues, planning multi-file refactors, evaluating tradeoffs, or producing outputs where mistakes are expensive (security review notes, migration plans, or incident postmortems). Opus 4.6 also supports adaptive thinking, where the model decides when deeper reasoning is warranted, and you can control how selective it is using an “effort” setting. The practical takeaway: extended thinking is best used when you want the model to spend more compute to reduce error rates, especially on tasks that involve hidden constraints or multiple interacting components.
In engineering workflows, the best heuristic is “enable it when you would normally reach for a whiteboard.” For example, if you ask, “Refactor this module without breaking public APIs and update all call sites,” that’s a planning-heavy task: the model should inspect types/interfaces, map dependencies, then propose a sequence of safe changes. Another good case is “explain why this bug happens,” where it must connect log evidence, code paths, and edge cases. In contrast, if the task is straightforward (“generate a serializer,” “write a simple unit test”), extended thinking is usually unnecessary and can slow down response time. A practical pattern is to expose two modes in your product: Fast (default) and Deep (extended thinking), and route requests automatically based on complexity (file count touched, presence of stack traces, length of retrieved context, etc.).
Extended thinking pairs well with retrieval because grounded context reduces the “search space” the model must reason over. If you use Milvus or Zilliz Cloud to retrieve the exact spec sections, API contracts, and constraints relevant to the task, then “deep thinking” becomes “deep reasoning on the right evidence,” which is where it pays off most. If you enable extended thinking without grounding, the model may spend extra compute reasoning about assumptions that aren’t true for your system. So the best practice is: retrieve authoritative context first, then enable extended thinking when the answer must be correct and defensible.