OpenAI's GPT-5.1 is an obsessive researcher, and Cline gives it focus (cline.bot)

OpenAI’s GPT-5.1—available in Cline now as GPT-5.1 and GPT-5.1-codex—behaves like an “obsessive researcher”: instead of grabbing minimal context, it maps entire codebase landscapes, produces exhaustive, coherent summaries, and can drift into unnecessary exploration unless guided. Cline’s team adapted agent prompts, tightened tool specifications, and sharpened a Plan vs Act split so the model channels that depth productively. Key operational tweaks include a Focus Chain (a persistent todo list that returns to context every six turns) and a persona that enforces iterative, execution-focused behavior; GPT-5.1 often needs only one concise example to internalize new rules. Technically, this model shines when planning is separated from execution: using /deep-planning or Plan Mode, GPT-5.1 generates architectural blueprints with exact file paths, function signatures, and stepwise execution sequences; Act Mode then reliably implements them. Cline expanded workflows into five phases—silent reading, silent terminal investigation, focused clarification, a scoped implementation plan, and a self-contained execution task—preventing premature tool calls and enabling richer, more grounded plans. GPT-5.1 offers a 272K context window (GPT-5.1-codex: 400K, coding-optimized). Pricing is $1.25 per million input tokens and $10 per million output tokens. For long-running, multi-phase engineering, this combination of deep research and structured scaffolding unlocks ambitious, reliable automation of complex development work.