Claude guide maps models to developer workloads

The useful takeaway is that model selection is becoming a workload-design problem, not a simple capability ladder.

• –Opus is positioned as the default for demanding engineering and agentic work, while Fable handles major refactors, migrations, and deep research
• –Sonnet fits scoped coding and repeated agent tasks where speed and cost matter
• –Haiku is the sensible choice for routing, extraction, classification, and subagents whose mistakes are inexpensive to catch
• –Adjusting effort before changing models could reduce both latency and spend
• –The guide gives developers a practical starting framework, but production evals should decide whether the cost-quality tradeoff holds for their workloads