Claude Fable 5.1 typically guesses quite than searches. Here is the repair

On this context, “Let me test quite than guess” is Claude’s approach of claiming it’s about to kick off an internet search quite than counting on its native reminiscence, which has a cutoff date that may be weeks and even months previous. Claude is being diligent about double-checking its details quite than assuming.

In my latest checks, most consumer-grade ChatGPT, Claude, and Gemini fashions are fairly diligent about looking quite than assuming, with even the puny-but-ultra-fast Claude Haiku 4.5 pausing to go looking earlier than sounding off on, say, Apple’s just-announced Mac mini fashions.

Surprisingly, Claude Fable 5.1—Anthropic’s strongest new Claude mannequin—does generally tend to skip the looking and simply guess, primarily when it’s set to low effort ranges. 

Phrase of that curious quirk comes straight from Anthropic itself, which notes in its Fable 5.1 prompting information that when its reasoning effort is about to “low,” the mannequin “is much less possible than Claude Fable 5 to name a search or retrieval instrument, and extra prone to reply from reminiscence.”

Anthropic’s first tip for making Fable 5.1 keep in mind to go looking is to easily increase its effort degree, holding in thoughts that you would be able to enhance Fable’s effort setting for only a single conversational “flip” (and thus eat fewer AI tokens) quite than for the complete chat.

There’s another choice, nevertheless: a “nudge towards verification,” which comes within the type of a Fable 5.1-specific system immediate: