5.5 Extended Thinking could not solve this one
On May 13, someone in that forum reported something worth noting: Claude 3.5's Extended Thinking mode had been given a problem, worked through it at deliberate length, and failed to solve it. The framing is what makes this interesting rather than trivial. Extended Thinking is not Claude running in standard conversational mode. It is Anthropic's architecture for sustained reasoning — the configuration that gives the model explicit space to slow down, surface its intermediate steps, and chain its logic toward an answer rather than pattern-matching toward one. When that mode fails, the failure carries different weight than a quick reply going wrong. It is not the model being hasty. It is the model trying its hardest, by design, and still coming up short. We do not know what the problem was. That detail, absent from the signal, matters enormously — a contrived adversarial puzzle and a genuine reasoning challenge are not the same thing, and without it, the observation floats without an anchor.
If confirmed, here is what this means. The AI industry has spent the better part of two years selling deliberate reasoning — chain-of-thought, extended thinking, deep research modes — as the architectural answer to the limitations of fast, reflexive generation. The implicit promise is that given enough compute and enough reasoning steps, these models can work their way through problems that would otherwise defeat them. A documented failure in Extended Thinking mode does not shatter that promise, but it quietly complicates it. It suggests there is a class of problems where more thinking time does not close the gap — where the issue is not computation depth but something more fundamental about the model's knowledge or its conceptual reach. For enterprise users building workflows around Extended Thinking as a reliable last resort, that is a meaningful constraint to map.
Watch for independent attempts to reproduce the failure with the same problem, and for any response from Anthropic that either acknowledges the limitation or clarifies what Extended Thinking is not designed to handle.
<table> <tr><td> <a href="https://www.reddit.com/r/ChatGPT/comments/1ur64xu/can_your_ai_solve_this/"> <img alt="Can your ai solve this?"
<table> <tr><td> <a href="https://www.reddit.com/r/ChatGPT/comments/1s5qnoa/so_i_asked_chatgpt_to_make_me_a_crossword_puzzle/"> <img alt="So I asked ChatGPT to make me a crossword
<table> <tr><td> <a href="https://www.reddit.com/r/ChatGPT/comments/1sywrqy/image_20_is_able_to_solve_geometry_problems_with/"> <img alt="Image 2.0 is able to solve geometry
<table> <tr><td> <a href="https://www.reddit.com/r/ChatGPT/comments/1s5k36b/has_anyone_seen_this_before_in_the_thinking/"> <img alt="Has anyone seen this before in the Thinking
<!-- SC_OFF --><div class="md"><p>I just got this pop-off saying that thinking mode now actually improves the accuracy of the images, and it does... like quite a lot. Suddenly, it
<table> <tr><td> <a href="https://www.reddit.com/r/ChatGPT/comments/1w0yy1r/why_is_it_so_close_yet_so_stubborn_and_dumb_at/"> <img alt="Why is it so close yet so stubborn and dumb
<table> <tr><td> <a href="https://www.reddit.com/r/singularity/comments/1tgn8qt/gemini_32_flash_is_capable_of_solving_imo_2025_p6/"> <img alt="Gemini 3.2 Flash is capable of
<table> <tr><td> <a href="https://www.reddit.com/r/OpenAI/comments/1whphi5/okay_i_think_i_finally_get_a2a_i_was_thinking/"> <img alt="Okay… I think I finally get A2A. I was
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →