For a long time, AI development was guided by the assumption that the more computing time a model is given, the better its answer will be – particularly for complex tasks. A new Anthropic study published in July 2025 now challenges this assumption. The researchers tested language models including Claude Sonnet 4, o3-mini and DeepSeek R1 across different task types and compared them with OpenAI models. Their conclusion: more computing time does not necessarily lead to better answers. In fact, models such as Claude Sonnet 4 and DeepSeek R1 often produced worse results when given more time to reason.
The researchers describe this effect as “Inverse Scaling in Test-Time Compute”, whereby models can become distracted by secondary issues, over-optimise or draw more incorrect conclusions as reasoning time increases. Performance declined particularly noticeably on deductive tasks.
For practical AI use, it is therefore worth considering carefully how much reasoning time a model should be given for analytical tasks – an important shift in perspective for companies.
By Ida Kandler
Source: t3n digital pioneers, 23 July 2025, by Noëlle Bölling
