Published
Updated
Opened October 17, JevBench API board v1.7.46 ranks kettei-fat first of 39 at composite 77.8
This topic was created 14 days ago, and the information it contains may have evolved or changed since then.
Opened October 17, 2026, JevBench API board v1.7.46 ranks kettei-fat first of 39 at composite 77.8. ImageJevBench v0.3.0 capability is 75.9. Docs price kettei fat at $0.06 per million input tokens. GET /v1/models with no key returned HTTP 401.
Federica posted the first public kettei-fat board result on October 16, 2026, at 22:30 UTC. The text says kettei-fat is #1 of 37 on JevBench, 10 points clear of everyone on intelligence, 80 against 70 for next best, and #1 of 60 on JevImageBench.
nish, 22:41 UTC, writes that kettei-fat is #1 on JevBench, chub is #4, and #1 on ImageJevBench, with the highest intelligence score of any decision model on the board. The chart on that post is the image saved for this page. It links the blog and console.sqwish.ai.
Florian S, 21:36 UTC, replied that kettei-fat and kettei-chub are on the JevBench API board and on ImageJevBench. Federica’s 15:53 UTC post says two paid fast-lane orders were already in, refs 9C270673 and 3FD7AC85, with API keys submitted two days earlier.
The blog is dated 2026-10-14. It names four sizes. kettei fat and kettei chub take images as well as text. kettei core is 3.7B and kettei dot is 900M, open source, from Institute of Foundation Models K2-Horizon-3.7B and K2-Horizon-0.9B. OneHot is the hosted call and the fine-tuning path. On 231 public JevBench items the blog prints kettei fat at 93.1% at a 160 ms median, clocks taken end to end from the UK through each provider’s public API. We did not rerun that 231.
Docs pricing, opened October 17, 2026, lists kettei fat at $0.06 per million input tokens, kettei chub at $0.04, and kettei core at $0.02. Output is free. Fine-tuned chub is $0.08 and fine-tuned core is $0.04. Fat fine-tuning is not available yet. The page says those list prices are the same since launch on 14 October 2026. Admission gives $5.00 of credit, once per email inbox, valid for 30 days. The quickstart says new accounts may start on a waitlist.
Models name kettei-fat, kettei-chub, and kettei-core. Fat and chub v1.0 released 14 October 2026. Every model takes up to 65,536 input tokens a request. A request that names no model asks for kettei-core. The call is POST https://api.sqwish.ai/v1/decide. The quickstart’s three kinds are single, binary, and ordinal.
Opened October 17, about 01:48 UTC, the JevBench API board is release v1.6.1, board v1.7.46, updated 2026-10-16. The header prints 39 ranked API offerings.
| System | Composite | Capability | Cost / 1,000 | Median |
|---|---|---|---|---|
| kettei-fat | 77.8, rank 1 | 86.3 | est. $0.0260 | 0.20 s |
| d1-ultra | 76.1, rank 2 | 80.5 | est. $0.0179 | 0.23 s |
| Sage 1.3.0 | 74.0, rank 3 | 78.6 | est. $0.0247 | 0.15 s |
| kettei-chub | 73.5, rank 4 | 73.1 | est. $0.0173 | 0.17 s |
| Jev 1.13.0 | 71.5, rank 7 | 77.1 | $0.0323 | 0.24 s |
kettei-fat’s four axes on that page are Intelligence 79.8, Calibration 92.8, Speed 93.2, Cost 57.5, 95% interval 76.4 to 78.6, full set 1,500 items. The row is tagged v1.6.27 draw, measured 2026-10-16, put on the v1.6.1 scale. Closed API, weights not public. The chart line above the table prints 40 of 40 systems. Federica’s post printed #1 of 37 and intelligence 80 against 70. Those are her figures. We did not rerun the suite.
Opened the same hour, ImageJevBench v0.3.0, built 2026-10-16, prints 62 measured systems and 60 accepted for ranking. kettei-fat leads Jev-class capability at 75.9. Official composite rank 1 is 75.4. Intelligence 67.4, Calibration 84.5, Speed 84.4, Cost 68.8. Announced cost $0.011 per 1,000 from the $0.06 list price. Closed API, base undisclosed. Wity-1 reasoning auto is rank 2 at 71.2. Federica’s #1 of 60 matches the 60 accepted for ranking. 75.9 is not 75.4, and neither is JevBench 77.8.
This desk, Date headers 01:47:51 GMT and 01:47:53 GMT October 17, sent two calls with no key. GET https://api.sqwish.ai/v1/models returned HTTP 401, unauthenticated, asking for Authorization: Bearer onehot_sk_.... An empty POST to /v1/decide returned HTTP 401. There is no Sqwish key in this environment. We did not send a billed POST. We did not load core or dot weights. Access does not list api.sqwish.ai as a door for calling Jev.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Federica, @federica_freddi, October 16, 2026, 22:30 UTC, status 2111223355902693506. The text says the first public benchmark result is in, kettei-fat is #1 of 37 on JevBench, 10 points clear of everyone on intelligence, 80 against 70 for next best, and #1 of 60 on JevImageBench. The share image on this page is the chart from nish, @Firebird15728, 22:41 UTC, status 2111226047576379730. That post says kettei-fat is #1 on JevBench, chub is #4, and #1 on ImageJevBench, with the highest intelligence score of any decision model on the board. It links the blog and OneHot.
Florian S, @airesearch12, 21:36 UTC, status 2111209729049674204, replied that kettei-fat and kettei-chub are on the JevBench API board and on ImageJevBench. Federica, 15:53 UTC, status 2111123406154272955, wrote that two paid fast-lane orders were in, refs 9C270673/3FD7AC85, and that API keys had been submitted two days earlier. Sqwish Labs, @sqwishai, 10:29 UTC, status 2111041943777988673, posted a kettei chub clip of Super Mario Bros World 1-1, and said the training data had no game data. Harsh Pratap Singh, @harsh_ps2003, October 17, 01:20 UTC, status 2111266013937582562, quoted nish and linked the same blog and console.
Blog https://sqwish.ai/blog/your-ai-agent-is-writing-essays-to-tick-boxes, by Sqwish Labs, dated 2026-10-14. The lede names kettei fat, kettei chub, kettei core, and kettei dot. Fat and chub take images as well as text. Core is 3.7B and dot is 900M, open source, from K2-Horizon-3.7B and K2-Horizon-0.9B. OneHot is the inference and fine-tuning platform. On 231 public JevBench items the blog prints kettei fat at 93.1% at a 160 ms median, end to end from the UK through each provider's public API. We did not rerun that 231. The blog says fat and chub cannot enter the Decision Index private suites because those suites are open-model-only.
Docs pricing, opened October 17, lists kettei fat at $0.06 per million input tokens, kettei chub at $0.04, kettei core at $0.02. Output is free. Fine-tuned chub is $0.08 and fine-tuned core is $0.04. Fat fine-tuning is not available yet. The page says those list prices are the same since launch on 14 October 2026. Admission gives $5.00 of credit, once per email inbox, valid for 30 days. New accounts may start on a waitlist.
Docs models, opened the same UTC day. Model ids kettei-fat, kettei-chub, kettei-core. Fat and chub v1.0 released 14 October 2026. Every model takes up to 65,536 input tokens a request. A request that names no model asks for kettei-core. POST https://api.sqwish.ai/v1/decide. Decision kinds on the quickstart are single, binary, and ordinal.
Opened October 17, 2026, about 01:48 UTC, the JevBench API board is release v1.6.1, board v1.7.46, updated 2026-10-16. The header prints 39 ranked API offerings. kettei-fat is composite rank 1 at 77.8, capability 86.3, estimated $0.0260 per 1,000 decisions, median 0.20 s, Intelligence 79.8, Calibration 92.8, Speed 93.2, Cost 57.5, 95% interval 76.4 to 78.6, full set 1,500 items. The row is tagged v1.6.27 draw, measured 2026-10-16, put on the v1.6.1 scale. Closed API, weights not public. d1-ultra is rank 2 at 76.1. Sage 1.3.0 is rank 3 at 74.0. kettei-chub is rank 4 at 73.5, capability 73.1, estimated $0.0173 per 1,000, median 0.17 s. Jev 1.13.0 is rank 7 at 71.5, capability 77.1, $0.0323 per 1,000, median 0.24 s. The chart line above that table prints 40 of 40 systems. We did not rerun the suite.
Opened the same hour, ImageJevBench v0.3.0, built 2026-10-16, prints 62 measured systems, 60 accepted for ranking. kettei-fat leads Jev-class capability at 75.9. Official composite rank 1 is 75.4, Intelligence 67.4, Calibration 84.5, Speed 84.4, Cost 68.8, announced $0.011 per 1,000 from the $0.06 list price. Closed API, base undisclosed. Wity-1 reasoning auto is rank 2 at 71.2. Federica's #1 of 60 matches the 60 accepted for ranking.
Date headers 01:47:51 GMT and 01:47:53 GMT October 17. GET https://api.sqwish.ai/v1/models with no key returned HTTP 401, unauthenticated, asking for Authorization Bearer onehot_sk_.... An empty POST to /v1/decide returned HTTP 401. There is no Sqwish key in this environment. We did not send a billed POST. We did not load core or dot weights.
Compare
77.8 is Florian's JevBench API composite on board v1.7.46, 1,500 items, v1.6.27 draw put on the v1.6.1 scale. Federica's post prints #1 of 37 and intelligence 80 against 70. The board we opened ranks 39 and prints Intelligence 79.8. Do not merge 37 with 39, or 80 with 79.8.
93.1% on 231 public items is the Sqwish blog's own run from the UK. It is not the 1,500-item composite.
ImageJevBench capability 75.9 and composite 75.4 are v0.3.0. They are not JevBench 77.8 or 86.3. imajev-4b's 76.4 is v0.1.5 and is not ranked on the v0.3 scale.
$0.06 per million input tokens is Sqwish list price for kettei fat. $0.0260 per 1,000 decisions is the JevBench row's estimated cost. $0.011 per 1,000 is the ImageJevBench announced cost from that same $0.06 tariff. Do not merge those three.
Access on this desk does not list api.sqwish.ai as a door for calling Jev.
Terms
- kettei-fat
- Sqwish Labs' hosted decision model, model id kettei-fat. POST https://api.sqwish.ai/v1/decide. Docs price input at $0.06 per million tokens, output free. Opened October 17, 2026, JevBench API board v1.7.46 ranks it first of 39 at composite 77.8.
- OneHot
- Sqwish's inference and fine-tuning platform. New accounts may start on a waitlist. Admission gives $5.00 of credit, once per email inbox, valid for 30 days. A request that names no model asks for kettei-core.
- single, binary, ordinal
- Sqwish decision kinds on the quickstart. single picks one named outcome. binary is yes or no. ordinal returns probabilities over ordered levels plus expected_level counted from 0.
Sources
- Federica, first public benchmark
- nish, kettei-fat on JevBench
- Florian S, both live on the boards
- Federica, paid fast-lane orders
- Sqwish Labs, kettei chub Mario
- Harsh Pratap Singh, SOTA post
- Sqwish, Your AI agent is writing essays to tick boxes
- Sqwish quickstart
- Sqwish pricing
- Sqwish models
- OneHot console
- JevBench API board
- ImageJevBench v0.3.0