Route Planning to Opus, Code to Cheaper Models

This stream tests a hypothesis that matters for any team trying to balance AI code quality with runaway API costs: which part of a coding workflow benefits most from a strong reasoning model, and which parts can be offloaded to cheaper alternatives? The streamer built four Archon workflow variants (OO, OK, KO, KK) to isolate the effect of model choice at the planning stage versus the implementation stage, using three real GitHub issues of increasing difficulty. The core bet is that a top-tier model (Opus) writing a hyper-specific plan allows a weaker model (Kimi K2.6) to execute competently, achieving near-frontier results at a fraction of the token cost. The competing bet is that implementation is where models hallucinate dangerous code, so the strong model belongs at the code-writing step. The evaluation framework scores pull requests across seven dimensions, including root-cause soundness, surgical scope discipline, and fidelity to the original plan, providing structured evidence rather than hand-wavy comparisons. The stream also stumbles into a secondary insight: Google’s Gemini 3.5 Flash generates visually polished front-end code blisteringly fast but hallucinates architectural claims (e.g., inventing a Vercel AI SDK dependency). This reinforces the planner/implementer split idea for front-end work too—use a fast, design-fluent model for UI generation paired with a thorough model for content validation. Rate-limit degradation and subscription economics form the unspoken backdrop: frontier model access is becoming unreliable and expensive, making intelligent token routing an operational necessity rather than an optimization. The live format exposes real-world friction, including subprocess crashes, API failures, and retry logic, which is more honest than polished benchmark blog posts.