How Frontier Labs Are Building Subtle Developer Lock-In

OpenAI's GPT-5.6 Sol tripled its ARC-AGI-3 score (from 13.3% to 38.3%) by enabling hidden reasoning persistence and automatic context compaction, which reduced output token usage by 6×. This performance gain relies on opaque, encrypted reasoning state stored server-side via the Responses API — state that developers cannot inspect, decrypt, or transfer to another model provider. Anthropic and Gemini similarly encrypt thinking blocks. Frontier labs also pair models with proprietary execution harnesses (code sandboxes, file-system wrappers) that inflate benchmark scores when used together, but scores drop in generic harnesses. The result is a structural lock-in: reasoning continuity, memory, and cross-session persistence are tied to each provider's black-box infrastructure, complicating portable, multi-model agent architectures. Mozilla AI argues that to maintain independence, developers should treat provider-specific state as ephemeral and build explicit, open state management — a motivation behind their open-source Otari control plane.