“According to feedback from developers, smaller language models can be less cost-effective than larger models like Anthropic's Opus for certain tasks because they require more tokens and time, or fail to complete the task at all.”