AI coding model-role framework: runner, deep thinker, agent, UI-first¶
Insight: AI coding in 2026 operates in distinct layers requiring different models: fast/cheap "runners" for constant Q&A (Claude Haiku 4.5, Gemini Flash 3), deep thinkers for architecture/debugging (Claude Opus 4.5, GPT 5.2 Codex), agentic loops for issue-to-patch workflows, and UI-first work where visual verification matters. The model layer alone isn't sufficient — the product/runtime layer (IDE, CLI, visual editor) determines how effectively you reach the model's ceiling.
Detail: Builder.io frames this as model + product = stack. The same model behaves differently across products because the execution layer (context packing, tool loops, output format) shapes results. Claude Opus 4.5 at $5/$25 per million tokens is positioned as dramatically cheaper than GPT-5.2 Pro at $21/$168, making it viable for deeper tasks. For frontend/UI work specifically, visual verification in the product layer reduces design drift — positioning tools with live rendering above pure-text IDEs.
Sources
Related: existing entry "Claude Code excels with structured context" in external/claude-code.md — COMPLEMENTS