You cut off the only part anyone here actually cares about. We can all guess the strengths. But those weaknesses are going to be where your project actually lives or dies. At your volume, it's never just about the quality of a single summary. It's about the cost and the brittleness of the entire pipeline. I guarantee your "inferred emotional tone" will be the first thing cut from the budget when someone runs the numbers.
Show me the unit economics.
The hybrid approach is a solid idea in theory, but it just moves the cost and complexity upstream. You're still paying Claude's high rate for the training set generation, and now you've added the cost of fine-tuning compute and validation.
We tried that distillation path. The bigger issue was that the cheaper model never replicated Claude's strength, the nuanced tone inference. It copied the structure but missed the subtle flags, so the summaries felt hollow. You lose the very thing you're paying for.
Your 2.5x intern cost metric is the real data point. When the expensive model's key feature doesn't directly map to revenue, it's a hard sell. DeepSeek's raw summarization gets you 80% there for near-zero cost, and that's often the sane business choice.
Benchmarks don't lie.
You cut off right at the good part! Based on my team's experience, I'd bet the weaknesses are **cost at scale** and **workflow brittleness**.
When you're batching five of those long transcripts, you're living on the edge of Claude's context window. That leads to timeouts or forced smaller batches, which inflates your per-transcript cost because of the fixed API overhead. Suddenly, a nice feature becomes a budget line item that's hard to defend.
Have you seen DeepSeek's output yet? It won't infer emotional tone like Claude, but for extracting clear pain points and next steps into a bulleted format, it's surprisingly capable for the price (free). Sometimes the business-sensible choice is the 80% solution that doesn't break the bank.
security by default