Claude Opus 4.6 FAQ Summary: Can’t Find the Entry Point, Long-Conversation Truncation, and Reasons for Refusals

Many people using Claude Opus 4.6 run into issues like “can’t select the model,” “it forgets earlier context mid-chat,” or “it suddenly refuses to answer.” Below is a FAQ organized by the most common scenarios, aiming to explain the typical pitfalls of Claude Opus 4.6 in a practical, actionable way. After reading, you should be able to tell whether it’s a settings issue, a prompting/writing issue, or the content itself triggering restrictions.

Can’t find the model entry point: Where do you select Claude Opus 4.6?

Claude Opus 4.6 is usually selected from the “Model” dropdown on the new chat page. If you don’t see it in the current conversation, start a new chat and look again. For some accounts, when the service is under heavy load or access hasn’t been enabled, Claude Opus 4.6 may not appear in the list—switching network environments or refreshing the page is often more effective than repeatedly clicking buttons.

If you’re using it within a Project, check whether the project’s default model is locked to another option. After switching the default model to Claude Opus 4.6, only new conversations will follow that setting. Existing old conversations generally won’t automatically “switch brains”; you’ll need to reselect Claude Opus 4.6 at the top of the chat or start a new thread.

Long conversations “forget earlier context”: How to reduce information loss with Claude Opus 4.6?

As a conversation gets longer, Claude Opus 4.6 will compress the context, which shows up as gradually losing details and earlier constraints no longer holding. The most reliable approach is to write key requirements as a “fixed instruction” block and place it in the Project instructions area or at the start of each conversation, listing non-negotiable rules in a checklist.

When you need to carry over complex background, it’s recommended to first have Claude Opus 4.6 restate the currently confirmed assumptions in bullet points, then use that restatement as the opening “baseline draft” of a new conversation. For long texts or lengthy requirements, paste the material in segments and have Claude Opus 4.6 output a “recorded summary” after each segment—this can significantly reduce the chance of it drifting off track.

Sudden refusals / overly conservative output: How to handle refusals in Claude Opus 4.6?

When Claude Opus 4.6 refuses, it’s usually because your request touches a safety boundary, or the way you describe it causes the system to misinterpret your intent. Prioritize two things: rewrite your goal as a compliant use case (e.g., “defense/compliance/education”), and clearly state that you don’t need executable details—only principles, risk points, and alternative approaches.

If you’re sure the content itself is compliant, try narrowing the task scope: first ask Claude Opus 4.6 what parts are allowed to be discussed, then refine step by step. Often it’s not that it can’t answer, but that you asked for too much all at once; switching to a step-by-step questioning style is more likely to pass.

Results don’t follow requirements: Language, format, citations, and “it looks made up”

Claude Opus 4.6 will occasionally mix English into Chinese, or have inconsistent formatting. The fix is to hard-spec the output requirements, such as: “Simplified Chinese throughout, use level-2 headings, no more than three sentences per paragraph, conclusion first then justification.” For code/table outputs, explicitly specifying “output Markdown only / output JSON only and do not explain” saves time compared with asking it to revise afterward.

As for citations, Claude Opus 4.6 won’t magically know which article you want to cite. If you want verifiable citations, it’s best to provide links, original excerpts, or file contents, and require it to “mark citation positions by paragraph.” When you feel it’s “plausibly filling in the gaps,” add a line like “If unknown, say unknown—don’t guess,” which usually reduces hallucinations.