A green Continue button scored noul 0.99 on jev-1.14.0, n = 3, and an empty field scored 0.02

Verify

Primary post is Diogo Almeida, @CompleteSkeptic, October 14, 2026, 21:52 UTC, status 2110488925823488497. The text says Jev was an image model the whole time, disabled because models tend to be less intelligent on other modalities. Header: TypeSafe-Danger-Zone-Images: true. The still on this page is the astronaut meme from that tweet.

Justin Schroeder, @jpschroeder, 21:07 UTC, status 2110477540863381920, wrote that TypeSafe announced image support in the Discord town hall, with the same images header and TypeSafe-Danger-Zone-Long-Context: true. The attached video is about 5 seconds.

Opened October 15, 2026, docs.typesafe.ai/images is titled Images. A Dreamer-ready callout says images and Box questions are not yet up to TypeSafe's production bar, that the request format and limits can change, and that each request with an image or a Box question must send TypeSafe-Danger-Zone-Images: true. An image can sit in state, instructions, or criteria. Jev 1.13 and Jev 1.14 both accept images. Box questions need Jev 1.14. Formats are PNG or JPEG, still images only. Animated PNGs are rejected. File size 16 MiB per image. Width and height 256 to 1568 pixels. Total pixels 78,400 to 2,458,624. A 256 by 256 image is too small. Request body 32 MiB. All images in the request 32 MiB after processing. Expanded request 32 MiB after every reference is replaced. The API does not resize a rejected image. EXIF orientation is not applied. Image tokens bill at the text input price. Audio and video stay out. The models table, opened the same day, still prints Input Text on both Jev rows.

The blog, opened the same day, says image support is fully composable via find-and-replace in assets, and that all models, Jev included, perform worse on multimodality than on pure text. Breaking questions into simpler decompositions helps.

On October 15, 2026, Date header 01:22:14 GMT, a POST to https://api.typesafe.ai/v1/systemone with an image in assets, model jev-preview, and no images header, returned HTTP 400. The message is Images are experimental. Send the header TypeSafe-Danger-Zone-Images: true to use them. Desk clock 410 ms.

Date header 01:22:15 GMT, the same picture with that header, model jev-preview, HTTP 200, named jev-1.14.0. The state is one 400 by 400 PNG drawn here: a red square from 0.40 to 0.80 of the width and 0.20 to 0.70 of the height on a light field. Noul Is there a red square in photo returned 0.39. The Box on that call is on the Box page. Usage 760 input tokens and 38 output tokens. Desk clock 457 ms. List-price arithmetic $0.000032.

The same second, model jev-latest, header on, Noul only, HTTP 200, named jev-1.13.0. Noul 0.90. Usage 746 input tokens and 21 output tokens. Desk clock 489 ms. List price $0.000031. Key, request JSON, and request ids are not recorded here. This desk did not send a photograph of a real object.

Date headers 06:16:36 GMT to 06:16:42 GMT, this desk sent the same 400 by 400 PNG three times with the images header. jev-1.14.0 Noul 0.49, 0.47, and 0.42, mean 0.46. Matching jev-1.13.0 Noul-only calls 0.94, 0.94, and 0.93, mean 0.94. An empty 400 by 400 field, no red square, scored 0.04 on jev-1.14.0 and 0.01 on jev-1.13.0. Date header 06:16:36 GMT, a POST with the present image and no header returned HTTP 400, Images are experimental. The Box answers from those three present calls are on the Box page. n = 3 on the present square. n = 1 on the empty field.

Wes Bos, @wesbos, posted at 23:15 UTC October 14, status 2110509898958008515, Jev images is here. One question about an 850 by 850 image billed $0.000096936, about 10,000 images for a dollar. The same post says referencing the image more than once duplicates token usage, so two questions in one request is equivalent to uploading two images. At 23:27 UTC, status 2110512898329948565, he wrote that the max is 1568 by 1568, and that 1568 is 4x more expensive than 784 by 784.

Opened October 15, the images docs say each use of a reference adds the image again, and that several questions that need the same file should put it in state once and refer by path. Image tokens bill at the text input price. The page says guidance on estimating those tokens will come later.

Date headers 11:04:54 GMT to 11:04:57 GMT, this desk sent six image POSTs with TypeSafe-Danger-Zone-Images: true, synthetic PNGs drawn here, a red square from 0.40 to 0.80 of the width and 0.20 to 0.70 of the height. Model jev-preview named jev-1.14.0. One Noul, image in state: 784 by 784 used 1,884 input tokens, list price $0.000079, noul 0.52, desk clock 449 ms. 850 by 850 used 2,238 tokens, $0.000094, noul 0.58, 442 ms. 1568 by 1568 used 6,589 tokens, $0.000277, noul 0.56, 711 ms. Two Nouls on the 850 file in state used 2,259 tokens. Two Nouls that each named the same 850 reference in instructions used 4,209 tokens, $0.000177. A matching 850 Noul on jev-latest named jev-1.13.0, 2,235 tokens, noul 0.90. n = 1 per layout. Key, request JSON, and request ids are not recorded here.

Egor Merkushev, @egormerkushev, posted at 13:59 UTC, status 2110732422614745368, Released Jev4Mellea 0.7.0, experimental image support for Jev 1.14: PNG/JPEG inputs for Noul, Choice, and Score. Opened later that UTC day, github.com/SoundBlaster/Jev4Mellea is Apache-2.0, tag v0.7.0 at 3b1bdf68a1e0, 13:52 UTC, 3 stars. The README says bounding boxes sit outside this adapter's image interface. This desk did not install the package.

The file examples/evaluation/visual/live-comparison-2026-10-15.md, opened the same day, records twelve procedural scenes on jev-1.14.0, started 11:45:44 UTC, source commit 9c5e03bf, adapter 0.6.0 from that checkout. Six expected accepts and six rejects. Labels from geometry. At reject 0.10 / accept 0.90: FAR 0 of 6, FRR 0 of 6, uncertain 7 of 12, coverage 5 of 12, accepted 0, rejected 5. At 0.20 / 0.80: coverage 9 of 12, accepted 3, rejected 6. Median 214.3 ms. The matching Luna route on CoreInfra is a different host. This desk did not rerun the 12.

brainstormity posted at 14:08 UTC, status 2110734677887574482, open-sourced JEVsnap, a screenshot redactor powered by Jev 1.14. Opened later, github.com/brainstormity/JEVsnap is MIT, public, v0.2, commit b249546b2538 at 05:54 UTC, 1 star. Default AI path sends OCR text with local secrets replaced by [HIDDEN], never the image. Box visual check is off by default. This desk did not install the app.

Date headers 14:57:24 GMT to 14:57:27 GMT, this desk sent a 400 by 400 PNG drawn here, a blue square from 0.08 to 0.28 of both axes and a red square from 0.40 to 0.80 of the width and 0.20 to 0.70 of the height. Three jev-preview calls named jev-1.14.0. Blue Noul 0.98, 0.98, 0.98. Red Noul 0.55, 0.55, 0.57, mean 0.56. Usage 797 to 798 input tokens. Desk clocks 5,455 ms, 476 ms, 778 ms. Matching jev-latest Noul-only calls named jev-1.13.0, blue 1.0, 1.0, 1.0, red 0.88, 0.88, 0.84, mean 0.87. The Box answers sit on the Box page. n = 3. Key, request JSON, and request ids are not recorded here.

Kerry Lu and Mateo Wang's LiteLLM post, dated October 15, 2026, adds day-0 image support on POST /v1/systemone and POST /v1/decisions. Client headers that start with TypeSafe-Danger-Zone- go to TypeSafe only. Images and box questions ship in the next LiteLLM release. Cost tracking for jev-1.14.0 works after a cost-map reload. This desk did not run the proxy.

Date headers 00:28:47 GMT to 00:28:48 GMT October 16, 2026, three jev-preview calls named jev-1.14.0 on a 400 by 400 Checkout PNG drawn here. Noul Is there a green Continue button in photo returned 0.98, 0.99, 0.99, mean 0.99. Matching jev-latest Noul-only call, Date header 00:28:48 GMT, named jev-1.13.0 and scored 1.0. Date header 00:28:49 GMT, a POST with that image and no images header returned HTTP 400.

Date header 00:28:50 GMT, four Nouls on a four-color PNG, image in state, all scored 1.0. Date headers 00:28:52 GMT to 00:28:53 GMT, three Nouls on a red circle among eighty gray circles scored 0.99, 0.99, 0.98. Date header 00:28:53 GMT, an empty field scored 0.02. The Box answers sit on the Box page.

Compare

imajev-4b is an Apache-2.0 LoRA that reads up to two photos. Its Image JevBench v0.3.0 capability is 68.0. That board is not this header. Clef-omni's October 9 post prices a 21-second video with sound at about 1.5 seconds. Jev's image docs reject video.

OpenAI's Decisions guide, on the OpenAI Jev page, takes inline base64 images. That host is OpenAI. The 400 message above is TypeSafe's.

0.90 and 0.39 are two n = 1 Nouls on a synthetic red square. They are not an image-board ranking. 0.39 is the 1.14.0 answer. 0.90 is the 1.13.0 answer. The Box edges from the 1.14.0 call are on /news/jev-box/.

The 06:16 UTC three-call means, 0.46 on 1.14.0 and 0.94 on 1.13.0, are a later sample of the same PNG. Do not merge 0.46 with 0.39. The empty-field pair 0.04 and 0.01 is n = 1.

Wes Bos billed $0.000096936 on an 850 by 850 photo. This desk's 850 PNG used 2,238 tokens, list price $0.000094 at $0.042 per million input. Those are different files. 6,589 / 1,884 is 3.50. Wes Bos printed 4x for 1568 against 784. Two questions that refer to an image in state added 21 tokens. Two questions that each wrote the reference into instructions used 4,209 tokens. The 11:04 UTC 850 Noul 0.58 is not the 0.46 mean on the 400 by 400 PNG.

Jev4Mellea's 5 of 12 coverage is twelve geometry-labeled scenes, Noul thresholds 0.10 / 0.90, on jev-1.14.0. It is not this desk's red-square 0.46 and not Image JevBench 68.0. Luna's 11 of 12 on that file is a different host.

The 14:57 UTC two-square PNG is a later sample. Blue 0.98 and red 0.56 on 1.14.0 sit next to a blue square that was not in the 06:16 PNG. Do not merge 0.56 with 0.46. JEVsnap is a wrapper. This desk did not run it.

The 00:28 UTC Continue noul 0.99 is a later sample on a button drawn here. Do not merge 0.99 with blue 0.98 or with Aviz's 82 percent. Empty-field 0.02 is n = 1 on a new PNG, not the 06:16 empty 0.04. Four-color Nouls 1.0 are n = 1. LiteLLM's docs illustration noul 0.99 is a worked example.