AI tip to cut your token costs by literally 80%: (works in Claude, Codex - literally any harness/model) Barbell strategy + subagents. Here's the idea: Most people run every part of a task through the same model. Let's say they use Fable 5 for planning, execution, and review. This is probably the most wasteful way to use these frontier models right now. I've been using a barbell approach that splits the work by intelligence needs. Heavy on both ends, light in the middle → Planning (10%): use your smartest, most expensive model to architect the task, define the spec, and set the success criteria. This is where intelligence actually matters - use top-tier models like Fable here. → Execution (80%): Tell your frontier model to spin up cheaper subagents to do the actual grunt work, writing, coding, research, based on the plan already laid out. → Verification (10%): bring the expensive model back in at the end to check the output against the original plan to catch errors, and confirm it's actually done right. I've shared this tip before but wanted to rehash it as it works exceptionally well with Sol + Luna Max.(Miles Deutscher)