Abe's 1,774 Snowflake posts cost $0.055 on Jev Noul against $1.87 on AI_FILTER

ABE Masanori posted on October 24, 2026, at 11:20 UTC. On 1,774 of his 2025 X posts, Jev Noul cost about $0.055 against AI_FILTER at about $1.87, about 1/34. AUC against AI_FILTER was 0.9685. Of 70 splits where Jev scored under 0.2 and AI_FILTER was TRUE, he judged Jev right on 54. Date headers 08:49:12 GMT to 08:49:40 GMT October 26, this desk's jev-1.14.0 noul on his Lightdash example was 0.04.

ABE Masanori posted on October 24, 2026, at 11:20 UTC. The text names a Qiita write-up of Snowflake AI_FILTER against Jev Noul on 1,774 posts, cost about 1/34. At 10:07 UTC the same day he wrote that a UDF to Jev looked slow until it hit the 80 requests per second rate limit.

The article is dated 2026-10-24. The run is 2026-10-20, AWS us-west-2, Standard edition, X-Small Gen2 warehouse. The rows are his 2025 X posts that are not reposts, 1,774 from a 28,642-row archive. The question asks whether the subject is a DBMS product itself. BI tools, ETL, catalogs, and product-free SQL theory are marked no.

Jev went through a Python UDF with external access. The request is POST https://api.typesafe.ai/v1/systemone, model jev-latest, response jev-1.13.0. AI_FILTER ran in SQL. He did not label the full set.

Cost, clocks, and splits

The printed total is about $0.055 for Jev, API $0.0486 plus warehouse $0.0061, against about $1.865 for AI_FILTER. Jev billed 1,158,233 input tokens. The Jev query ran 23.0 s at about 80 requests per second, median 53 ms, max 560 ms. AI_FILTER ran 7.1 s. Errors 0.

Treating AI_FILTER TRUE as the positive class, AUC is 0.9685. AI_FILTER said TRUE on 324 of 1,774, 18.3%. Jev’s median score is 0.02, and 77% of scores sit under 0.1.

He read the splits by hand. Of 70 posts where Jev scored under 0.2 and AI_FILTER was TRUE, he judged Jev right on 54, AI_FILTER on 5, and 11 mixed. Of 14 where Jev scored 0.5 or above and AI_FILTER was FALSE, 5 Jev, 3 AI_FILTER, 6 mixed. A 20-row repeat shifted scores by at most 0.07. Snowflake Decision through AI_COMPLETE was private preview on his account and stayed out.

One published example in that 37-row low-score group is a Lightdash yml note, Jev 0.03, AI_FILTER TRUE. He judged no.

This desk, October 26

Date headers 08:49:11 GMT, 08:49:12 GMT, and 08:49:40 GMT October 26. GET https://api.typesafe.ai/v1/models returned HTTP 200 listing jev-latest and jev-preview. A POST with model jev-latest named jev-1.14.0. His Lightdash sentence, noul, 0.04, 525 input tokens, 22 output, desk clock 415 ms. jev-preview named jev-1.14.0, noul 0.04, desk 409 ms. A named jev-1.13.0 POST named jev-1.13.0, noul 0.04, desk 421 ms.

0.04 is this desk on jev-1.14.0. Abe’s 0.03 on that sentence is jev-1.13.0 in the 1,774. This desk did not rerun the 1,774.

Jevflake is Luke Kranz’s September 20 dbt package of SQL functions. Abe’s UDF is a different call from the same warehouse product. Neither is a catalog row.

This site's reading

Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.

Verify

ABE Masanori, @abe_masanori, October 24, 2026, 11:20 UTC, status 2113953846267703306. The text names a Qiita comparison of Snowflake AI_FILTER and Jev Noul on 1,774 posts, cost 1/34. The 10:07 UTC post, status 2113935347478659513, says a UDF to Jev looked slow until it hit the 80 requests per second rate limit.

The Qiita article, posted 2026-10-24. Data: 1,774 non-repost posts from his 2025 X archive, run 2026-10-20 on an X-Small Gen2 warehouse in AWS us-west-2, Standard edition. Question: whether the post's subject is a DBMS product itself, with BI, ETL, catalogs, and product-free SQL theory marked no. Jev via a Python UDF, POST https://api.typesafe.ai/v1/systemone, model jev-latest, response jev-1.13.0. AI_FILTER via SQL. No gold labels on the full set.

Printed table: Jev about $0.055 (API $0.0486 plus warehouse $0.0061) against AI_FILTER about $1.865. 1,158,233 Jev input tokens. Jev 23.0 s, about 80 requests per second, median 53 ms. AI_FILTER 7.1 s. AUC 0.9685 treating AI_FILTER TRUE as positive. AI_FILTER TRUE 324 of 1,774 (18.3%). Jev median 0.02, 77% under 0.1. Of 70 posts where Jev scored under 0.2 and AI_FILTER was TRUE, he judged Jev right on 54, AI_FILTER on 5, and 11 mixed. Of 14 where Jev scored 0.5 or above and AI_FILTER was FALSE, 5 Jev, 3 AI_FILTER, 6 mixed. Errors 0. A 20-row repeat shifted scores by at most 0.07. Snowflake Decision through AI_COMPLETE was private preview on his account and stayed out.

Date headers 08:49:11 GMT, 08:49:12 GMT, and 08:49:40 GMT October 26. GET https://api.typesafe.ai/v1/models returned HTTP 200 listing jev-latest and jev-preview. A POST with model jev-latest named jev-1.14.0. His published Lightdash sentence, noul, 0.04, 525 input tokens, 22 output, desk clock 415 ms. jev-preview named jev-1.14.0, noul 0.04, desk 409 ms. A named jev-1.13.0 POST named jev-1.13.0, noul 0.04, desk 421 ms. Abe's article prints 0.03 on that sentence for jev-1.13.0. n = 1 each. This desk did not rerun the 1,774.

Compare

$0.055 against $1.865 is Abe's 1,774-row total at published list prices, about 1/34. Token unit prices in that article are $0.042 per million Jev input against $3.24 per million AI_FILTER tokens. Do not merge 1/34 with 1/77.

0.9685 is AUC against AI_FILTER, not accuracy. 54 of 70 is his hand read of one split, not a labeled F1. 23.0 s is the Jev query at 80 requests per second. 7.1 s is AI_FILTER. 53 ms is his UDF median. This desk's 415 ms is one later call from another host.

0.04 is this desk's jev-1.14.0 noul on the Lightdash example. Abe's 0.03 on that sentence is jev-1.13.0 in the 1,774. Do not merge those two. Jevflake's September 20 SQL functions are a different package. Snowflake Decision stayed out of his table.

Terms

AI_FILTER: Snowflake Cortex function that classifies a prompt as TRUE or FALSE. Abe compared it with Jev Noul on 1,774 of his 2025 posts.
1/34: Abe's cost ratio on those 1,774 posts: Jev about $0.055 against AI_FILTER about $1.87, at published list prices. The token-unit ratio in that article is about 1/77.