Screenshots from AI leakerCHOIhave dropped on go today , showing a new “ extended thought ” toggle and a web search tool on the Claude roving app . There ’s no information yet about how these features will work or when they ’re fare , but it depend like the extended thinking lineament will be forPro usersonly .
Claude Extended Thinking!pic.twitter.com / DAImsEYPrt
& mdash ; CHOI ( @arrakis_ai)February 19 , 2025
Claude is one of the lesser - known generative AI assistants out there , always losing the spotlight to ChatGPT , Gemini , and now DeepSeek as well . That does n’t discontinue it from following the latest LLM movement , however . Judging by the name and the little clock image next to it , “ extended thinking ” will plausibly be an in advance reasoning feature that will let Claude drop longer cerebrate about its answers and show its view process .
This could be similar toOpenAI ’s o1and o1 - miniskirt models , which “ think ” for longer and include some “ reasoning ” in their answers . It could also be similar toDeepSeek R1 ’s chain - of - thoughtfeature that let in a footfall - by - stone’s throw account of the fashion model ’s opinion process separate to its last result . Gemini ’s Flash Thinking framework also shows a elaborated breakdown of the “ intellection process ” used to reach an answer or render a response .
The primary difference between the advanced reasoning feature of these competing models is how much of the thought process we see . OpenAI , for example , pick out to hide the strand of thought of its o1 model so it could be left completely unfiltered and used for monitoring function . No one has any idea what kind of content will bug out up in a wholly unedited rendering of an LLM ’s “ thought process , ” so the company clearly does n’t want to show it to users .
By that logic , the range - of - thought account that DeepSeek R1 andGemini instant Thinkingshow are likely edit out or controlled in some way to prevent too many strange “ thoughts ” popping up . We do n’t know which way yet Claude will incline with its lengthened thinking feature article but the ultimate end is likely the same — to take more time and bring forth more “ reasoned ” responses to complex prompts .
As for the connection search beta that shows on the second screenshot , directly accessing the connection is something Claude users have been wanting for a while . Until now , the good example has been designed to be “ ego - contained , ” but with so many other company addingweb - surfing abilitiesto their models , it seems Anthropic has deemed it a necessary addition to stay competitive .
It ’s unclear right now who will get access to these raw features and when , but I go for we ’ll all get the chance to fiddle around with them at least a small spot .
