At 00:28 UTC October 16, TypeSafe's jev-latest still returned jev-1.13.0, and Aviz Maeir scored Jev 1.14 Box at 82% on 180 scenes

Verify

Primary post is Diogo Almeida, @CompleteSkeptic, October 14, 2026, 21:52 UTC, status 2110488924309348860, Happy JevDay. The thread says the first Jev release favored stable infrastructure, and this one is for early adopters, with a gate into a danger zone. Later tweets in that thread cover images, a Box type, longer context, an opt-in to help find jaggedness, jev-1.14.0 as jev-preview, and three internal charts. The blog link is status 2110488935269077041. TypeSafe's teaser, status 2110419371453825349, 17:16 UTC, asked what's this and pointed at a town hall with Almeida and @allietheicon at 1:14 PM. Almeida's 17:24 UTC reply named 1:14 PM PST and Discord. 1:14 PM PDT on October 14 is 20:14 UTC.

Opened October 15, 2026, typesafe.ai/blog/jev-1-14 is titled Jev: For the builders AND dreamers, dated October 14, 2026. The TL;DR is Jev 1.14.0 plus experimental images, a new type, and longer context. jev-preview is the new version. Largest gain expected in multilingual support. Model releases should generally be strict improvements. jev-latest will point at jev-1.14.0 in a week. jev-1.13.0 stays available. TypeSafe intends 60 days warning before jev-1.13.0 is formally deprecated, and 30 days for jev-1.14.0, and says a deprecation policy will come first.

The models page, opened October 15, lists Jev 1.14 and Jev 1.13 in one table. Both rows print $42 per billion tokens and $0.042 per million, 64k per-request and 32k per-question, and Input Text. Jev 1.14 is available from October 14, 2026. Output tokens are free. Aliases: jev-latest points at jev-1.13.0, the most recent stable official release. jev-preview points at jev-1.14.0. GET /v1/models lists the aliases. Versioned ids such as jev-1.14.0 are accepted whether or not they appear in that list.

The three charts in Almeida's evals tweet, and again in TypeSafe's 01:02 UTC October 15 thread, are watermarked antibench.com and carry the box Warning: AI benchmarks Please don't trust them. Overall intelligence: jev 1.14.0 60.0, jev 1.13.0 58.9, kev-27b 46.7, pplx-decider-v1.1-27b 46.0, clef 44.3, clef-flash 38.8, kev-4b 36.9, d1-3b 33.8. Overall calibration error, lower better: 1.14.0 1.6, 1.13.0 1.8, kev-4b 10.3, clef 11.2, kev-27b 17.8, clef-flash 19.1, d1-3b 19.9, pplx-decider-v1.1-27b 35.0. The intelligence versus cost scatter puts both Jev points near $100 per million queries on that axis. Almeida wrote that the measure is an experiment, likely biased toward TypeSafe's models, and that hosted models were left out of the comparison because of their data policies. The charts still print Jev in pink. This desk did not rerun that suite.

On October 15, 2026, Date header 01:22:12 GMT, GET https://api.typesafe.ai/v1/models with a TypeSafe key returned HTTP 200. The list had two names: jev-latest, description The latest iteration of TypeSafe's System One Model: Jev, release_date 2026-09-10T18:38:01.391457+00:00; jev-preview, description A preview version of jev-latest: should be better in most ways, release_date 2026-09-10T18:39:06.057655+00:00. Desk clock 445 ms.

Date header 01:22:13 GMT, POST /v1/systemone, model jev-latest, state Help! My payouts have been failing for 3 days., Noul is_urgent and Choice department. HTTP 200, model jev-1.13.0. Noul 0.95. Choice billing, probabilities 0.88, 0.12, 0.0, confidence 0.82. Usage 361 input tokens and 57 output tokens. Desk clock 427 ms. List-price arithmetic $0.000015. A second POST the same second, model jev-preview, HTTP 200, model jev-1.14.0. Noul 0.95. Choice billing, 0.86, 0.14, 0.0, confidence 0.79. Same token counts. Desk clock 433 ms. Date header 01:22:14 GMT, model jev-1.14.0 by id, same answers as jev-preview, desk clock 654 ms. Key, request JSON, and request ids are not recorded here.

Later the same UTC day, Date headers 03:03:53 GMT to 03:03:56 GMT, six POSTs with model jev-preview returned HTTP 200 and named jev-1.14.0. Four states are the docs' payouts sentence in English, Chinese, Spanish, and Japanese. The questions stayed in English: urgency Noul, department Choice, frustration Score, and a four-way language Choice. Language came back at 1.0 on the matching label for each of the four. Department chose billing on all four: English 0.86 at confidence 0.79, Chinese 0.99 at 0.98, Spanish 0.96 at 0.94, Japanese 0.99 at 0.98. Urgency was 0.95, 0.96, 0.93, and 0.94. Frustration on English and Spanish scored 1.02 at confidence 0.97. Japanese scored 1.10 at 0.85. Chinese scored 1.37 at 0.44, with mass 0.63 on Frustrated and 0.37 on Very angry. Desk clocks 466, 429, 425, and 427 ms. Tokens 504/118, 511/119, 505/119, and 513/120.

A fifth state, Date header 03:03:55 GMT, is a Chinese API-500 login outage written here. Department chose technical at 1.0. Urgency 0.85. Language chinese at 1.0. Tokens 513/119. Desk clock 431 ms. A sixth POST, Date header 03:03:56 GMT, reused the Chinese payouts sentence with questions written in Chinese. Department stayed billing at 0.99, confidence 0.98. Urgency 0.93. Frustration scored 1.33 at confidence 0.50. Tokens 543/119. Desk clock 437 ms.

Five matching POSTs with model jev-latest returned jev-1.13.0, Date headers 03:03:56 GMT to 03:04:52 GMT. Language was also 1.0 on English, Chinese, Spanish, and Japanese. Billing stayed the top department on the four payouts tickets. The Chinese outage also chose technical at 1.0. Frustration on Chinese was 1.54 at confidence 0.31, with mass 0.46 on Frustrated and 0.54 on Very angry. Japanese frustration was 1.25 at 0.62. Eleven calls, 5,635 input tokens and 1,309 output tokens. List-price arithmetic at $0.042 per million input tokens, output free, is $0.000237. Key, request JSON, and request ids are not recorded here.

OpenRouter posted Jev 1.14 is now on OpenRouter at 00:16 UTC, status 2110525059093987468. Opened later that UTC day, the model page https://openrouter.ai/typesafe/jev-1.14 is titled TypeSafe: Jev 1.14, id typesafe/jev-1.14, In / Out Price $0.042 / $0 per 1M, Context 262K, Released Oct 14, 2026. The provider table lists TypeSafe, input $0.042, output Free, P50 latency 0.15s, uptime 100.00%. The hub https://openrouter.ai/docs/guides/community/jev names typesafe/jev-1.14 and says ~typesafe/jev-latest tracks the newest release. Opened the same day, https://openrouter.ai/~typesafe/jev-latest is titled TypeSafe: Jev Latest. The heading links Jev 1.14. The page says this model always redirects to the latest model in the Jev family. GET https://openrouter.ai/api/v1/models did not list typesafe/jev-1.14. GET of that id and of ~typesafe/jev-latest returned HTTP 404. This desk did not call OpenRouter. No OPENROUTER_API_KEY is in the environment.

At 04:12 UTC Almeida replied that the antibench warning is not humor, that TypeSafe does not want people to blindly trust those evals, and that the chart is a best guess and is biased.

Date headers 06:16:20 GMT to 06:16:44 GMT, this desk sent 32 POSTs to https://api.typesafe.ai/v1/systemone, plus GET /v1/models. GET listed jev-latest and jev-preview. Every jev-preview POST named jev-1.14.0. Every jev-latest POST named jev-1.13.0. Three Chinese payouts tickets, English questions, jev-1.14.0 frustration 1.27 at 0.59, 1.32 at 0.52, and 1.32 at 0.52, mean 1.30 at mean confidence 0.54. Matching jev-1.13.0 scores 1.55 at 0.32, 1.52 at 0.28, and 1.55 at 0.32, mean 1.54 at mean confidence 0.31. Language was 1.0 on English, Chinese, Spanish, Japanese, French, German, and Korean on both models. Department chose billing on every payouts ticket and technical at 1.0 on the Chinese API-500 outage. Thirty-two POSTs, 20,518 input tokens and 2,806 output tokens. List-price arithmetic $0.000862. n = 3 on Chinese, n = 1 on the other six languages. Images and Box from this sweep are on the linked pages. Key, request JSON, and request ids are not recorded here.

Göksu Güneş, @seoilan, posted at 10:02 UTC, status 2110672594810577078, that both models scored 29 of 30. The article, sosyaldijital.com/jev-1-14-test-sonuclari-dogruluk-musteri-kaybi/, opened the same day, says SosyalDijital sent 30 synthetic Turkish messages, three labels, buyer, seller, and other, one pass each, 60 API calls. Both models got 29 of 30. The shared miss is "Reklam değil gerçek talep: ürün fotoğraflarını düzenleteceğim, ücret konuşalım." Expected buyer, both returned seller. Recorded response model fields typesafe/jev-1.13-20260917 and typesafe/jev-1.14-20261014. Mean clocks 0.3669 s on 1.13 and 0.3518 s on 1.14. Recorded cost $0.001129884 for 60 calls, $0.000564942 per model. The protocol required at least 29 correct and zero missed buyers. This desk did not rerun the 30.

Date headers 11:04:52 GMT to 11:05:00 GMT, this desk sent 14 POSTs plus GET /v1/models. GET listed jev-latest and jev-preview, both release_date fields still 2026-09-10. A POST with model jev-latest named jev-1.13.0, noul 0.89, billing 0.97 at confidence 0.95, 353 input tokens, desk clock 438 ms. A POST with model jev-preview named jev-1.14.0, noul 0.90, billing 0.93 at 0.89, same token counts, desk clock 660 ms. Opened the same hour, openrouter.ai/typesafe/jev-1.14 still returned HTTP 200. GET /api/v1/models still omitted that id. GET of that id still returned HTTP 404. This desk did not call OpenRouter.

Six POSTs used that published buyer sentence as state, a Choice written here, buyer, seller, or other. Date headers 11:04:57 GMT to 11:04:59 GMT, three jev-preview calls named jev-1.14.0 and chose seller at confidence 0.48, 0.43, and 0.52, seller probabilities 0.65, 0.62, and 0.68. Matching jev-latest calls named jev-1.13.0: seller at 0.41, buyer at 0.28, seller at 0.43. Each of those six used 392 input tokens and 38 output tokens. Image token counts from the same minute sit on the images page. Key, request JSON, and request ids are not recorded here.

Yuuki Yamashita, @yama3133, posted a Qiita write-up dated October 15, 2026, Jev 1.14.0で京都弁の判定がどう変わるか、1.13.0と並べて測ってみた. Sixty frozen Kyoto-ben phrases, six labels, ten each, plus ten later phrases. Criteria matched the public web app. Model ids pinned jev-1.13.0 and jev-1.14.0. Seventy items three times each, 420 calls, zero retries, about 510,000 input tokens, about $0.02, same UTC day. Three-run mean accuracy 0.817 on 1.13.0 and 0.833 on 1.14.0, runs 0.800 / 0.833 / 0.817 against 0.850 / 0.833 / 0.817. Macro F1 0.816 and 0.832. The extra ten phrases scored 0.70 on both. Majority vote matched on all 60 test items. One majority change sat in the extra ten: うちの子ぉも、ええ子にしてますえ, expected warning, 1.13.0 sarcastic_criticism, 1.14.0 social_nicety. McNemar p = 1.000. Median clocks 178 ms and 176 ms. This desk did not rerun the 60.

Cristian, @XCSme, posted at 12:21 UTC, status 2110707709532065961, Jev 1.14 seems to do worse than 1.13 on my tests, with a link to aibenchy.com/compare/typesafe-jev-1-13-none/typesafe-jev-1-14-none/. Opened later that UTC day, the page prints Jev 1.13 default score 2.6, rank 407, tests correct 6 of 12, attempt pass rate 26.1 percent, 0 flaky, average 698 ms. Jev 1.14 default score 2.3, rank 412, tests correct 4 of 12, 23.2 percent, 2 flaky, average 377 ms. Both rows 36 of 69 attempts, 36 runs, 67,092 input tokens, 18,927 output tokens, $0.003. Adapters supply choices. Each row supports 12 of 23 tests. Other tests are unsupported, not failed. Score uses the full suite. Last updated 2026-10-15. This desk did not rerun that suite.

Ishir Lakhani, @IshirLakhani, posted at 14:16 UTC, status 2110736672136798536, a smash test: full game state, how high a CPU the model can beat in a best of 20. jev 1.14 got to level 3. No hardcoded behavior. Every button is the model's call. The attached video is about 45 seconds. No table. No repo. This desk did not rerun it.

Date headers 14:57:05 GMT to 14:57:28 GMT, this desk sent 33 POSTs to https://api.typesafe.ai/v1/systemone plus GET /v1/models. GET listed jev-latest and jev-preview, both release_date fields still 2026-09-10. A POST with model jev-latest named jev-1.13.0, noul 0.89, billing 1.0 at confidence 1.0, 366 input tokens, desk clock 501 ms. A POST with model jev-preview named jev-1.14.0, noul 0.90, billing 0.99 at 0.98, same token counts, desk clock 459 ms. Opened that hour, openrouter.ai/typesafe/jev-1.14 still returned HTTP 200, title Jev 1.14. GET /api/v1/models still omitted that id. GET of that id still returned HTTP 404. This desk did not call OpenRouter.

Eighteen POSTs used a six-way Choice written here on three Kyoto-ben phrases. Criteria names matched Yamashita's labels. The descriptions were written here. Three jev-preview calls named jev-1.14.0 on うちの子ぉも、ええ子にしてますえ and chose social_nicety at confidence 0.69, 0.66, and 0.64, warning mass 0.03, 0.04, and 0.04. Matching jev-latest calls named jev-1.13.0 and chose social_nicety at 0.33, 0.36, and 0.34, warning mass 0.09, 0.07, and 0.10. ぶぶ漬けでもどうどす? chose indirect_refusal on all six, confidence 0.94, 0.93, 0.93 on 1.14.0 and 0.87, 0.88, 0.84 on 1.13.0. 明日の会議は十時どす chose literal on all six. Each of those eighteen used 465 to 472 input tokens.

Six POSTs used a Japanese payouts sentence written here. Language chose japanese at 1.0 on all six. Department chose billing at 1.0 on all six. Urgency noul 0.84, 0.84, 0.83 on 1.14.0 and 0.89, 0.89, 0.89 on 1.13.0. Each used 462 input tokens. Image and Box from this minute sit on the linked pages. Thirty-two POSTs printed usage, 16,666 input tokens and 2,397 output tokens. List-price arithmetic $0.000700. One Box POST on jev-1.13.0 returned HTTP 400 and printed no usage. Key, request JSON, and request ids are not recorded here.

Images, Box, and 262k context are on the linked pages. The still on this page is TypeSafe's teaser art, What? JEV is evolving!

Aviz Maeir, @aviz85, posted at 22:04 UTC, status 2110854390999306324. Same 180 scenes against OpenAI Responses API plus strict JSON: Jev 1.14 82 percent, 0.21 s, $0.15 per 1k. gpt-6-luna 84 percent, 1.5 s, $0.22 per 1k. gpt-6.1-sol 95 percent, 3.4 s, $5.03 per 1k. Confidence 0.7 or above was right 107 of 108. The dated numbers sit on the Box page. This desk did not rerun the 180.

PC Watch posted at 22:19 UTC, status 2110858020553170981. The article, opened October 16, is dated 07:18 JST October 16, which is 22:18 UTC October 15, by 稲津 定晃. It reprints the blog: jev-preview, $0.042, danger-zone images, box on 1.14.0, 262k, jev-latest moves in a week. No independent table.

Kerry Lu and Mateo Wang's LiteLLM post, dated October 15, adds day-0 support for Jev 1.14 images, box, and long context. This desk did not run the proxy.

Date headers 00:28:46 GMT to 00:28:53 GMT October 16, 2026, this desk sent 16 POSTs plus GET /v1/models. Date header 00:28:46 GMT, GET returned HTTP 200. Date header 00:29:34 GMT, a second GET named jev-latest and jev-preview, both release_date still 2026-09-10. A POST with model jev-latest named jev-1.13.0, noul 0.89, billing 0.87 at confidence 0.80, 359 input tokens, desk clock 411 ms. A POST with model jev-preview named jev-1.14.0, noul 0.90, billing 0.86 at 0.79, same token counts, desk clock 396 ms. Fourteen POSTs printed usage, 24,552 input tokens and 674 output tokens. List-price arithmetic $0.001031. Image and Box from this minute sit on the linked pages.

Opened that hour in the connected Chrome tab, openrouter.ai/typesafe/jev-1.14 titled Jev 1.14, In / Out Price $0.042 / $0 per 1M, Context 262K, Released Oct 14, 2026. TypeSafe P50 0.19 s, uptime 100.00 percent. GET https://openrouter.ai/api/v1/models returned HTTP 200 and did not list typesafe/jev-1.14. GET of that id returned HTTP 404. An anonymous HTML GET of the page returned HTTP 403. This desk did not call OpenRouter.

Taka Sasaki, @TakaSasakiDev, October 19, 2026, 04:26 UTC, status 2112037552844935448, wrote that Jev 1.14 is on Cloudflare AI. Opened that UTC morning, developers.cloudflare.com/ai/models/typesafe/jev-1.14.0/ returned HTTP 200. Model id typesafe/jev-1.14.0. Context window 32,000 tokens. Input $0.042 per million tokens, cached input $0.00, output $0.00. The page lists Noul, Choice, Score, and Box. Image inputs and Box questions are beta. Sample noul 0.95 on a black-rectangle PNG. We did not call Workers AI.

Opened October 19, about 07:20 UTC, JevBench JSON revision v1.6.29 ranks Jev 1.14.0 ninth of 38 hosted systems at composite 70.2, Capability 75.01, Intelligence 59.94, Calibration 90.08, p50 0.224 s, estimated $0.03175 per 1,000. measured_on 2026-10-16, measured_in v1.6.27. Jev 1.13.0 stays seventh at 71.49, Capability 77.13. Date header 07:24:57 GMT, this desk's jev-latest still named jev-1.13.0. We did not rerun the suite.

Compare

60.0 and 58.9 are TypeSafe's internal intelligence chart. They are not JevBench's hosted composite. Opened October 19, 2026, JevBench ranks Jev 1.14.0 ninth of 38 at composite 70.2, Capability 75.01. Jev 1.13.0 stays seventh at 71.49, Capability 77.13. Opened October 11, 2026, that board ranks Jev 1.13.0 fourth at 71.5. Decision Index 0.3's Jev reference of 60.1 is a third figure. Do not merge those three.

pplx-decider-v1.1-27b prints 46.0 on the internal chart and 62.8 as Decision Index 0.3 rank 1. Clef's 44.3 is not the October 1 median latency row. d1-3b's 33.8 is not open-d1's Decision Index 0.2.1 cell of 48.57.

This desk's payouts calls are n = 1 each. The September 30 three-question payouts POST also named jev-1.13.0 and also returned noul 0.95 and billing 0.88 at confidence 0.82. That older call included a Score and used 426 input tokens. The October 15 pair dropped the Score. 427 ms and 433 ms are this desk's clocks, not JevBench p50.

The 03:03 UTC four-language sweep is also n = 1 per ticket. Language 1.0 on jev-1.14.0 is also 1.0 on jev-1.13.0 for these four sentences. It is not the blog's multilingual-gain claim. The Chinese Score is the split: 1.14.0 1.37 at 0.44, 1.13.0 1.54 at 0.31. The 01:22 English two-question pair stays a separate sample. 466 ms through 640 ms are this desk's clocks.

The 06:16 UTC sweep is a later sample. Chinese n = 3 mean 1.30 at 0.54 on 1.14.0 against 1.54 at 0.31 on 1.13.0. Do not merge those means with the 03:03 singles. Language 1.0 on French, German, and Korean used a seven-option Choice. OpenRouter's 262K card is that host's listed window. TypeSafe's default table still prints 64k and 32k. The OpenRouter latest alias already linked Jev 1.14. TypeSafe's jev-latest still returned jev-1.13.0.

SosyalDijital's 29 of 30 is 30 synthetic Turkish messages through OpenRouter, one pass, with that site's unpublished prompt. This desk's n = 3 is one published sentence, a Choice written here, on TypeSafe's API. Do not merge 29 of 30 with language 1.0 or with Chinese 1.30. The 11:04 UTC alias pair is a later sample than 01:22 and 06:16. noul 0.89 and 0.90 on the English payouts sentence are not the earlier 0.95 pair.

Yamashita's 0.833 and 0.817 are 60 frozen Kyoto-ben phrases, three runs, unpublished criteria. This desk's n = 3 used a Choice written here on three phrases. Do not merge 0.833 with language 1.0 or with Chinese 1.30. Do not merge this desk's social_nicety on うちの子ぉも with Yamashita's majority pair sarcastic_criticism / social_nicety.

AI BENCHY 2.6 and 2.3 are that site's private suite with adapters supplying choices. Each row supports 12 of 23 tests. Score uses the full suite. It is not JevBench 71.5 and not TypeSafe's 60.0 chart. 377 ms and 698 ms are that page's averages.

The 14:57 UTC alias pair is a later sample than 11:04. noul 0.89 and 0.90, billing 1.0 and 0.99, stay separate from 11:04 0.89 / 0.97 and from 01:22 0.95 / 0.88. Japanese payouts language 1.0, n = 3, is the same language question as the 03:03 singles. Smash test level 3 is one tweet.

The 00:28 UTC alias pair is a later sample than 14:57. noul 0.89 and 0.90, billing 0.87 and 0.86, stay separate from 14:57 0.89 / 1.0. OpenRouter's TypeSafe P50 0.19 s is that host's listed clock on October 16. The October 15 reading printed 0.15 s. Aviz 82 percent is a Box test. Do not merge it with 0.833 or with 2.3. PC Watch reprints the blog.