hybrid-opus-haiku-model-allocation-for-researchclaw-23-stages
AutoResearchClaw’s 23 pipeline stages do not all require the same LLM quality. The decision was Opus 4.6 for 8 critical reasoning stages (hypothesis generation, paper writing, peer review, claim verification) and Haiku 4.5 for 15 routine stages (topic init, resource planning, export, formatting). Both models are routed via ACP through Claude Code, staying within the Anthropic ecosystem and the existing Max subscription. AutoResearchClaw natively supports primary + fallback model configuration.