Too expensive ($60 per million output tokens vs. $7.5)
No support for structured outputs (can't validate)
Could re-explore this with vanilla JSON mode (no structured outputs) and then fallback to the existing implementation if it doesn't go through, to compare the quality of insights.
Test using
o1-preview
for runtime lead analysis, with a fallback to 4o. Or, to Sonnet 3.5 with structured output to reduce provider dependency.