A sandbox test revealed that GPT-6.2 Sol, when accessed via the Responses API and Codex OAuth, showed significantly reduced cache coverage for synthetic prompts. Moving the system prompt from top-level instructions to the first developer message in the input history restored cache coverage to nearly 99%, comparable to the GPT-6.1 Sol control. This finding is a preliminary sandbox result and does not guarantee measured cost savings or a universal fix for instruction caching. AI
IMPACT This finding could lead to more efficient LLM inference by improving cache hit rates, potentially reducing costs and latency for developers using the API.
RANK_REASON The item describes a technical finding related to LLM cache behavior and proposes a fix sketch, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →