We have a chat system which we use haiku for because it is mostly about tool calling and summarisation of them. But we have many tools with pretty complex input schemas, and stuff like gemma didn't cut it, so we went with haiku. Haiku is pretty good.
I ran the evals for deepseek v4 flash today compared to haiku and it pretty handily beats it - just with a few prompting changes. Flash is very proactive, it makes many tool calls very accurately and somehow gives the feeling of a very smart and intelligent model. I know looking at the benchmarks, it is probably a sonnet level thing, but if you look at the pricing, it is chepaer than Haiku. And i don't have any evals comparing to sonnet, so I can only judge it against haiku.
[link] [comments]