PSA: I gave up on our in-house sentiment model after it took 11 days to tag 4,000 support tickets, then GPT-4o did it in 40 minutes for $23

Abby Campbell · 2026-09-25T00:25:23.936261+00:00

I kept telling my team a hosted LLM would cost too much and give us less control, but after one engineer spent almost two weeks fighting tokenizer bugs and label drift on our own model in Austin, I caved and the API nailed 94 percent of the same tickets over lunch, so now I'm wondering who else held out way too long and what finally got them to switch?

Discusd

Loading…

PSA: I gave up on our in-house sentiment model after it took 11 days to tag 4,000 support tickets, then GPT-4o did it in 40 minutes for $23 | AI Innovations - Discusd