But if your use case is entirely, “here’s the full context of details, analyze / summarize / extract” - it’s extremely reliable. I think you can see a lot of product releases recently that emphasize this exact use case.

Source: Lessons after a half-billion GPT tokens - Ken Kantzer’s Blog