GPT-5.6 (Luna, Terra, Sol) ships with a claim aimed straight at your bill: frontier coding scores on half the output tokens. What that means for agent economics, why you should verify it on your own workload — and how to build so model choice stays a config change.