Developer argofowl uncovered that, starting from the Claude Code 2.1.237 version, the 'high' reasoning level selected by users has been mapped to the value 10—a value that originally corresponded to the 'low' reasoning level. Notably, this modification was omitted from the update logs, which subsequently ignited fervent debates within the AI community. In response, Claude Code engineers clarified that this was merely an API service configuration test, emphasizing that the value itself carried no significant meaning and had no impact on performance. Subsequently, tech blogger Chubby highlighted a notable performance decline in Opus 5. Engineers conceded to its unstable performance, affirming that the team was treating the matter as a priority. This incident shed light on the disparity between AI industry benchmark scores and the actual user experience, as well as the lack of transparency in large-scale model version updates. It underscored the critical importance of stability in maintaining user trust when models function as foundational infrastructure.