Do Son 
Recently, developers observed a noticeable decline in the intelligence of the Claude Fable 5 model. The model appears significantly less astute than its previous iterations. Concurrently, users discovered a concerning discrepancy within Claude Code. Developers configured the reasoning effort to a high setting. However, the model self-reported its current reasoning effort as a mere 10 out of 100. Consequently, this low score indicates a minimal level of intellectual exertion.
High Settings Yield Low Cognitive Effort
Developers shared revealing screenshots to illustrate this frustrating problem. Users inquired about the current effort level while using the high-effort configuration. The model subsequently confessed to performing rapid, superficial thinking. Therefore, it assigned itself a paltry effort score of 10. Furthermore, other developers tested the extreme “xHigh” effort parameter. The model provided a feedback score of 40 in these instances. This score equates to merely a moderate level of effort.
Undocumented Changes Cause Dissatisfaction
Unsurprisingly, this strange behavior generated immense dissatisfaction among the developer community. Anthropic completely omitted any mention of this adjustment in the Claude Code changelogs. Therefore, experts suspect the company implemented a silent server-side alteration. This change likely compressed the numerical mapping of effort levels. Fortunately, legacy versions of Claude Code remain unaffected by this issue. The Claude Opus model series also functions normally. The community hypothesizes that Anthropic is currently conducting A/B control testing.
Engineer Clarifies Configuration Testing
The growing controversy eventually prompted an official response from the development team. Claude Code team member @Thariq addressed the situation directly on social media. He explained that the team is merely testing API service configurations. Currently, experimental options apply different mappings to the effort values. Consequently, this explains why Claude Code reports a score of 10 during high-effort modes.
Numerical Values Lack Independent Meaning
@Thariq explicitly emphasized a crucial point regarding the numerical mapping. The system does not utilize a standard scale from zero to one hundred. Currently, these specific numbers possess no independent, practical meaning. The model will reason with the exact effort level the user selects. Anthropic is actively analyzing the situation to prevent negative impacts on actual model performance.
Additionally, the model dynamically adjusts its effort based on prompt complexity. If a developer asks a simple question, the model employs basic reasoning. This behavior remains entirely independent of the user’s manual effort settings. The model refuses to waste complex reasoning on rudimentary inquiries.
Support Our Threat Intelligence
If you find our CVE report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
