Making Fable Cheaper Than Opus (cognition.com)

In a recent analysis, Fable 5 has outperformed Opus 4.8 while proving to be more cost-efficient when paired with a sidekick model in coding tasks. Although Fable 5 has a token cost that is double that of Opus, its innovative delegation strategy using the new Fusion architecture allows it to achieve superior results at a significantly lower overall cost. The experiment conducted on FrontierCode 1.1 demonstrated that a combination of Fable 5 and a cheaper sidekick model ended up costing $1.86 per run, compared to $2.04 for Opus 4.8 with its sidekick, with Fable also delivering a higher performance score (60.7 vs. 54.6). This shift underscores the significance of effective delegation in AI models, revealing that how an agent manages tasks can significantly impact costs and outcomes. Fable’s approach allows it to delegate the bulk of the work to its sidekick efficiently, rarely needing to engage in coding tasks itself, while Opus tends to take on more of the workload, thereby incurring higher costs. As sidekick models become more affordable and capable, the implications for the AI/ML community are profound; it emphasizes the importance of strategic delegation over mere high-quality outputs, potentially paving the way for more efficient and cost-effective AI-driven coding solutions.