base_model = "anthropic/claude-sonnet-4-5-20250929"
# Chat uses `thinking = { type = "enabled", budget_tokens = N }`; N >= 1024
# and must be less than `max_tokens`.
# https://docs.databricks.com/aws/en/machine-learning/model-serving/query-reason-models (accessed 2026-06-25)
reasoning_options = [{ type = "budget_tokens", min = 1_024 }]

[cost]
input = 3
output = 15
cache_read = 0.3
cache_write = 3.75