Dropstone 1.5: Technical Report (blankline.org)

Dropstone 1.5 has been announced as a versioned, model-agnostic coding runtime that integrates leading open-weight models into a secure and cost-effective coding CLI. This update marks a significant advancement in the AI/ML community as it formalizes a monthly evaluation cycle—the "eval cycle"—for selecting the best models to compose the runtime, ensuring continuous improvement based on real-world coding performance. The current composition includes DeepSeek V4 Flash for the Fast tier, DeepSeek V4 Pro for the Pro tier, and Moonshot Kimi K2.6 for the Heavy tier, each evaluated against specific agentic coding tasks. Technical details reveal that users can execute approximately 450 heavy coding turns weekly on a $15 monthly plan due to a highly efficient prefix-cache setup, achieving over 95% hit rates as sessions progress. This pricing model and cache strategy underscore Dropstone's focus on delivering predictable, affordable services without the uncertainties of traditional metered APIs. Notably, Dropstone emphasizes safety and compliance, as it operates entirely on US-hosted infrastructure, addressing data privacy concerns common with global models. Overall, Dropstone 1.5 presents a robust alternative for developers, combining superior performance, continuous model evaluation, and enhanced security features within a user-friendly package.