Where each shines
- Claude — long-context deal reviews, careful drafting
- GPT — voice, real-time, broad tool use
- Gemini — multimodal (call video, product screenshots)
- Open-source — cost-sensitive, on-prem enterprise
What matters more than model
Retrieval quality, prompt structure, guardrails, and eval infrastructure. Model choice is 20% of quality; the rest is engineering.
How often to re-evaluate
Quarterly. The frontier moves fast enough that a static choice ages in months.