Jev Just Got Eyes. It Beat GPT-6.1: 9× Faster, 20× Cheaper (1,000 Photos)

Author: Hyperautomation Labs · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Jev Just Got Eyes. It Beat GPT-6.1: 9× Faster, 20× Cheaper (1,000 Photos) · Channel: Hyperautomation Labs · 10:57 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI evaluation and benchmarks, Prompt and context engineering, Multimodal models

Summary

The video tests Jev, a low-cost AI model that returns structured answers rather than writing text, after it gains the ability to analyze images. Across six use cases, it performs well on refund claims, receipt detection, visual matching, and large-scale photo classification, while revealing a key failure: when asked to locate something, it may always pick a location even when the item is absent. The presenter recommends first checking whether the object exists, then asking where it is, and using Jev’s confidence to flag uncertain results for human review. In a 1,000-photo test, Jev was reported to be much faster and cheaper than GPT-6.1 Soul, with practical constraints including supported image formats and size limits.

Key points

Jev answers with probabilities or choices rather than free-form text, and the video presents its image analysis as inexpensive.
A first image test failed because the image was silently dropped by the route; the presenter recommends using the newest AI SDK through Vercel AI Gateway and checking token counts.
On 24 refund-claim examples, Jev got 23 right, and processed all 24 in under four seconds for a third of a cent.
When asked where a cat was in 20 photos without cats, Jev picked a grid square every time; the suggested safeguard is to ask whether the cat is present first.
For helmet detection, Jev identified whether riders wore helmets in 90 of 100 photos, but locating the relevant rider was harder; analyzing cropped quarters improved localization.
Jev correctly identified all 50 receipts and rejected all 15 non-receipt images, but could point to the total rather than read it.
In a visual matching test, Jev selected the matching breed photo 30 out of 30 times, with average confidence of 97%.
On 1,000 photos, Jev classified whether an animal was present with 994 correct results in 15 seconds at a reported cost of four cents; GPT-6.1 Soul took over two minutes and cost 86 cents.
The presenter says Jev’s uncertainty can help route cases to a human: most errors in the 1,000-photo test appeared among the 27 photos where Jev was unsure.
The video notes that Jev supports PNG and JPEG photos, not video; image sides must be between 280 and about 1,500 pixels, and phone photos should be rotated before submission.

Chapters

Start	Chapter	Summary
00:00	Jev gains vision and the first test	The presenter introduces Jev’s image-analysis capability, tests a screenshot, and discovers the image was silently dropped because of the route used.
02:05	Image setup and refund claims	After explaining how to send images and check token counts, the presenter tests Jev on refund claims and reports 23 correct answers out of 24.
04:11	The ghost problem and locating objects	Jev always picks a location when asked where a cat is, so the presenter recommends asking whether the object exists first and explores detection on busy street photos.
06:16	Receipts, visual matching, and scale test	The video covers receipt detection and visual matching, then compares Jev with GPT-6.1 Soul on a 1,000-photo classification task.
08:52	Practical ideas and limitations	The presenter suggests six applications, lists image constraints, and closes with advice to test Jev and use uncertainty as a warning.

Quotes

If your token count barely moves when you add a photo, your AI is blind. — 02:05

The fix is one extra question. Before you ask where, ask whether. — 04:41

You don’t need a perfect AI. You need one that knows when to ask you. — 08:52

Just never trust the weather without asking whether. — 10:27

Concepts and entities

Jev, GPT-6.1 Soul, Vercel AI Gateway, AI SDK, TypeSafe, box, grade 3×3, Hyperautomation Labs

Full transcript (original language)

00:00 A customer says their milkshake spilled all over the bag. They send a photo. And before you’ve even opened it, an AI has already looked at it and decided, “This photo backs up the story.” And this one, for the exact same complaint, it doesn’t. Two years ago, building that took a team of engineers and thousands of labeled photos. Now, it takes one sentence and costs a fraction of a cent.

00:31 And the strangest part, the AI doing it can’t write a single word. It’s called Jev. If you watched our first video, you know it only ever answers with odds. Yes or no. This or that. That’s why it’s so cheap. This week, it learned to see. So, I gave it six cases and made it play detective with over a thousand real photos. It solved one case perfectly. 30 out of 30.

01:02 And in another one, it pointed at something that wasn’t there 20 times in a row. If you ever build with this, that mistake could cost you real money. So, stay for case two. But my very first test went wrong in the most interesting way. I sent Jev a screenshot and asked two questions. Is there a screen in this photo? And is there a cat? It answered both in a third of a second. 38% and 24%.

01:34 Both low, both vague. That’s not a judgment. That’s a shrug. Then I looked at the bill. 303 tokens. A photo that size should cost about four times that. Jev never saw it. The route I used quietly dropped the image and still sent back an answer. No error, no warning. So, here’s tip number one before anything else. If you call Jev through Vercel’s AI

02:05 Gateway, update to the newest AI SDK and send the photo as a file. Then check your token count. Same photo, sent the new way. Nearly 1,300 tokens. And now the screen comes back at 97% and the cat at five. If your token count barely moves when you add a photo, your AI is blind. Here’s how it works. You hand Jev a photo as the evidence. You ask a question

02:36 and you give it the possible answers. It can say yes or no with a percentage. Pick one option from a list or give a score. TypeSafe also launched a brand new answer type called box where it draws a rectangle around whatever you describe. It’s labeled experimental and it isn’t on the gateway I’m using yet. So, for every where is it question, I did the next best thing. I cut each photo into a 3x3 grid and made Jev pick a square. Case one, the

03:07 refund. I wrote 24 refund claims and made a test photo for each one. Some photos back up the story. Some show a perfectly fine meal and a few are from something else entirely like a house plant for a crushed noodle box. For each one, Jev answers three things. Does the photo back up the claim? How bad is the damage? And where is the problem? It got 23 out of 24 right. All 24 checked in under 4 seconds for a

03:40 third of a cent. And the one it missed, it said 54%. A coin flip. Hold on to that. It matters at the end. If you run an online shop, a cloud kitchen, or anything with refunds, that’s your first pass checker. It doesn’t replace you. It just hands you the claims that actually need your eyes and asks everyone else for a better photo. Case two, the ghost. This is the one I promised you. I took 20 real photos with no cat anywhere.

04:11 A ski slope, a busy street, a clock, a pizza on a table. And I asked it one question. Where’s the cat? It never said there isn’t one. 20 photos, 20 answers, a square every single time. And this isn’t just my grid. Type safe’s own docs say the new box answer does the same thing. A box answer is always a box. There is no not found. Now picture that

04:41 on an insurance claim or a security camera. Confident and wrong. The fix is one extra question. Before you ask where, ask whether. Is there a cat in this photo? On all 20 empty photos, Jeff said no, every time. And on 20 photos that really did have a cat, it said yes, 20 out of 30. So the rule is simple. If the weather comes back low, throw the where away.

05:12 Screenshot this part. And if it just saved you from a nasty bug, hit like so more builders see it. Case three, the helmet patrol. A hundred real street photos of people on bikes and scooters. One question. Is every rider wearing a helmet? It got 90 out of a hundred right. And when someone wasn’t wearing one, it got 40 out of 41. But then I asked, which rider? And using the grid, it picked the right square only 11 times out of 41.

05:45 The heads are small and the photos are busy. So, I did what a detective does. I zoomed in. I cut every photo into four quarters and asked Jev about each quarter on its own. Is there a rider here without a helmet? Then, I picked the quarter it was most sure about. Asking about the whole photo, it found the right quarter 28 times out of 41. Asking each quarter on its own, 32. Better, not perfect. And five small questions still cost a

06:16 fraction of a cent. Same trick works on shelves, crowds, parking lots, anything big and busy. Case four, the suspicious receipt. If you file expenses or you’re the person approving them, this one’s for you. I gave it 50 real receipts and 15 photos that aren’t receipts at all. A laptop, a cake, a phone. Is this a real receipt? 50 out of 50, yes. 15 out of 15 fakes, rejected. Then,

06:48 where’s the total? It found the right square on 38 out of 47. One catch. It points, it doesn’t type. It can show you where the total is, but it can’t read the number back to you. So, pair it with any regular AI to read what’s inside that square. Case five, the one that made me sit up. This time, the answers aren’t words. They’re pictures. I showed Jev four photos of four different breeds with no names at all.

07:19 Just A, B, C, and D. Then, a brand new photo of one of them. Which one matches? I ran it 30 times with different breeds every time. 30 out of 30. Average confidence, 97%. And it works for anything you can photograph. Is this the same product as our catalog? Is this the right parcel? You show it what right looks like and it compares. One case is a demo. A thousand is a job.

07:49 So, I gave it the night shift. 1,000 real photos one simple question. Is there an animal in this photo? It finished all 1,000 in 15 seconds. 994 right. Total cost 4 cents. Then I ran the same thousand photos through GPT-6.1 Soul, one of the top chatbots. It took over 2 minutes and cost 86 cents. That’s nearly nine times slower and 20

08:21 times the price. And most of its misses, it counted people as animals, which to be fair we are. But speed isn’t the clever part. Remember that coin flip from case one? It tells you how sure it is. So, I split the thousand into two piles. When it was sure, which was 973 photos, it got all but one right. And the 27 photos where it was unsure, that’s where five of its six mistakes

08:52 were hiding. Its own doubt was the warning sign. You don’t need a perfect AI. You need one that knows when to ask you. So, what would you actually build with this? Here are six ideas and each one is a few lines of code. First, a refund checker for any online shop. Second, delivery proof. So, you know the parcel is at the door and it’s the right parcel. Third, a listing cleaner that rejects

09:23 blurry or fake product photos before they go live. Fourth, an expense sorter. Fifth, a second opinion for security cameras. So, a guard only checks the real alerts. And sixth, a photo library tagger that sorts 10,000 pictures overnight for pennies. Quick fine print. Photos only for now. PNG or JPEG. No video. Each side has to be between 280 and about 1,500 pixels or it gets rejected.

09:55 Rotate phone photos before you send them because the rotation tag gets ignored. And Type Safe itself calls this experimental. So, test it on your own photos before you trust it with money. Real talk. 90% of you watching haven’t subscribed. On Facebook, subscribe is the supporter button. That is what keeps the lights on here. On YouTube, it’s free and it tells the algorithm to send you more. Subscribe to support us. So, the AI that

10:27 couldn’t talk can now see. It’s fast, it’s almost free, and it knows when it’s unsure. Just never trust the weather without asking whether. I put all six build recipes, the ghost fix, and the size limits into a free starter kit. Comment the word witness and I’ll send it to you. And I want case seven to be yours. Tell me in the comments what photo would you make it judge? I’ll test the best one in a future video. If you know someone who handles refunds

10:57 or receipts all day, send them this. Follow along on Instagram and Facebook for bite-size AI. And the deep dives live here on YouTube.