High effort reasoning settings on Opus 5.5 generate heavy computing costs

A fresh round of updates from major AI labs has brought cheaper prices alongside some surprising performance quirks. Tech commentator Simon Willison reported that Anthropic released its upgraded Claude Opus 5.5 model, shortly followed by OpenAI launching its GPT-6 Sol and GPT-6 Luna models. While both companies reduced their per-token prices, testing revealed that setting Anthropic's new Opus model to its maximum thinking intensity led to extreme overthinking. In one instance, the AI spent nearly twenty minutes planning a vector graphic of a pelican riding a bicycle before running out of memory and failing completely.

For everyday users and developers, these findings show that turning on the highest reasoning settings can be a costly trap. Willison noted that each failed attempt at generating the pelican image cost over two dollars while yielding no final output at all. As tech providers push models that spend more time silently pondering complex prompts, users risk paying steep bills for background processing that might fail before finishing. On a brighter note, overall base pricing across competing platforms continues to fall, making standard text tasks cheaper than before.

It remains uncertain how model builders will balance deep problem solving against reasonable execution times. Will platforms introduce default compute caps to prevent runaway reasoning loops, or will customers need to monitor background costs manually? As AI systems take on longer planning processes, finding a balance between thoroughness and practical efficiency will be vital for keeping broad software tools affordable.