MotherDuck's prompt_jev() Brings AI Text Classification into SQL at 50x Speed, 1% Cost
MotherDuck released promptjev(), a SQL function powered by TypeSafe AI's new model Jev, enabling text classification directly inside DuckDB SQL. Official benchmarks show that on a 100K-row classification task, Jev finishes in 40 seconds for $0.5 with accuracy on par with (or better than) frontier LLMs, which would take 32+ minutes. Posts report roughly 50x the speed at about 1% of the cost, making previously impractical analysis workflows viable.
Confirmed
- MotherDuck shipped promptjev(), backed by TypeSafe AI's Jev, for in-database text classification.
- Official numbers: 100K rows in 40 seconds for $0.5, near-frontier accuracy; 50x speed at 1% cost.
- A developer released duckdb-jev, a DuckDB extension that asks questions about every row in SQL and returns true SQL types rather than text; data can stay in tables or Parquet files.
Why it matters
- Per comments relayed by @joshwills and others, this embeds AI classification where the data already lives, avoiding data movement and unlocking workflows previously too slow or expensive.
- Returning native SQL types instead of text means model outputs feed directly into downstream SQL computation, further cutting pipeline integration costs.
2026-09-22 ~ 2026-09-22 · 5 related posts
Primary sources
- Jev Lands in MotherDuck SQL: 50x Faster, 1% the Cost of Frontier Models — josh_wills · 2026-09-22
- [source] duckdb-jev: ask an LLM about every table row in SQL, get typed results back — dbreunig · 2026-09-22
- MotherDuck makes text classification 50x faster at ~1% of LLM cost — josh_wills · 2026-09-22
2 near-duplicate retellings: charles_irl · dbreunig