Text models
Route work across GPT-5.6 Sol, Terra, and Luna
Give each tier an acceptance threshold and compare the cost of usable responses.
Set a quality threshold per task
Include extraction, routine code changes, and ambiguous investigations. Validate the response format and accuracy before judging cost.
Start with a narrow routing rule
Apply the least expensive tier that passes a routine task. Define when it should escalate, and measure the impact on retries and human review.
Test prompt
Extract product, amount, and currency from the order record. Return only those fields and use null for missing values.