Disclosure: we make ThumbnailPeak, which has its own thumbnail generator, and we scored everything with ThumbnailPeak's own critique. Every image and score below is from 21 October 2026. The video is DIY PRO's, published under a Creative Commons licence.
Yes, ChatGPT can make a YouTube thumbnail, and it can look very good. We asked it for one three ways, for a real go-kart build video. Here's what happened:
- The first try scored 96 out of 100 in ThumbnailPeak's critique, higher than any of the four our own Generate made for the same video that day.
- But it shows the wrong go-kart. Asked with only the title, ChatGPT drew a red racing kart on a dirt track. The video is about a green home-built kart. A viewer who clicks for the red one won't find it.
- All three came back in the wrong shape. Each image was 1536×1024, a 3:2 rectangle. YouTube thumbnails are 16:9, so YouTube adds black bars at the sides, and cropping cuts off part of the picture.
- All three repeated the title. "$20 Rust Bucket" and "Kick $$$" are already in the title, and they were on every thumbnail too.
The good news is that most of this is fixable if you know to look for it.
How we tested
We picked DIY PRO's "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212": a rusty go-kart frame rebuilt into a bright green two-seater with a new Predator 212 engine. It's the video we've used in several earlier posts, so we know it well.
We didn't use the ChatGPT app. We used OpenAI's API in the way that comes closest to it: a chat model with OpenAI's image tool, set to chatgpt-image-latest, the image model the API offers under ChatGPT's name. The chat model reads your message, writes its own picture prompt and calls the image tool, as ChatGPT does. The app adds instructions of its own, so your results there may differ.
We sent three messages, one try each, and kept the first image every time:
- Title only. "Make a YouTube thumbnail for my video "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212"."
- Title and description. The same, with the video's description pasted below it.
- Title and frames. The same, with four frames from the video attached: the rusty frame, the frame painted green, the engine and the steering wheel.
Then we scored each one in ThumbnailPeak's AI critique with the video's title, as YouTube would show it: with black bars at the sides.
What ChatGPT got right
A lot, honestly.
- The idea. All three went for rust-to-finished, the same before-and-after DIY PRO used on the real thumbnail. That's the strongest idea for this video.
- Feed-size basics. Big subject, strong contrast, thick outlined text. The first try got the top grade on legibility, hierarchy, subject clarity, contrast and curiosity.
- Speed. Each took under a minute.
On the critique's numbers, the first try is the best thumbnail in this post. If all you need is a strong-looking picture for a video whose subject anyone could draw, ChatGPT can do that.
What it got wrong
It drew a kart that isn't in the video
With only the title to go on, ChatGPT made up a kart: a red-and-black racer with a nose cone, sliding on a dirt track. The prompt it wrote for the image tool asked for "vivid red frame accents" and "racing toward camera on a dirt track". DIY PRO's kart is bright green and has two seats.
Pasting the description didn't fix the colour: the second try drew a red two-seater, because the description says "2 seater" but never says green. Only the third try, with frames attached, drew a green kart. It still drew one seat.
Our critique scored the red kart 96 and never noticed, because it doesn't watch the video. It reads the picture and the title, and a red kart fits "$20 Rust Bucket into Kick $$$ Go Kart" perfectly well. That's worth knowing about any thumbnail score, ours included: it can tell you whether a thumbnail works with its title, not whether it's true to the footage. That part is your job.
It came back in the wrong shape
ChatGPT's own prompt asked the image tool for "16:9 landscape (1280x720 composition)". All three images came back 1536×1024, which is 3:2.
YouTube shows a thumbnail in a 16:9 frame. Upload a 3:2 image and it gets black bars at the sides. Crop it to 16:9 yourself and you lose a strip off the top and the bottom.
The fix is easy once you know: crop it to 16:9 with a thumbnail resizer and check that no words got cut, or ask ChatGPT to keep all text and the subject well inside the middle of the picture.
It repeated the title
Every try put the title's own words on the thumbnail: "$20 Rust Bucket" and "Kick $$$". The critique flagged all three for it. On the second it said the thumbnail "repeats nearly all of the title's main terms", and suggested a payoff like "FROM SCRAP TO SPEED" instead.
YouTube shows your title right under the thumbnail, so words that repeat it use up space that could say something new. It's an easy fix in a follow-up message: "don't use any words from the title".
The third try also changed "Kick $$$" to "KICK $$". Read every word before you upload.
What Generate did with the same link
For comparison, we pasted the same link into Generate, ThumbnailPeak's AI thumbnail generator, on the same day: once for a single concept, then for a batch of three. Generate draws with an OpenAI image model too. The difference is what happens before the drawing: it reads the video's frames, title and description itself, plans the idea, and asks for a 1280×720 picture.
They scored 89 to 92, below ChatGPT's 96. The single concept, "LIVE AXLE", was the weakest: the critique said it "shows separate parts rather than the claimed $20 transformation". But every one of the four shows DIY PRO's actual green frame and engine, and every one is already the right shape. Their words are short and mostly add something the title doesn't say, like "ROLLING AGAIN", though the critique said "FROM RUST TO RACE" "closely restates the title's central promise".
So it depends on what you need. ChatGPT made the best-looking picture. Generate made pictures of the video.
When to use which
ChatGPT is a good choice when:
- you already know exactly what the thumbnail should show, and can describe it;
- the subject is something anyone could draw, like a type of food, a place or a common product;
- you want to change it by chatting ("make the text bigger", "swap the background").
Watch for the shape, invented details and words from your title, and attach frames or photos of anything that has to look like the real thing.
Generate is built for when:
- the thumbnail has to show what's really in your video, and you'd rather paste a link than describe it;
- you want three different ideas side by side, not three versions of one;
- you want your own face in it from a face library, the same person every time;
- the video isn't filmed yet: you can describe it and get a thumbnail first.
It needs a free account, and the free plan includes 1 generation a month, which makes one concept.
If you use ChatGPT, do this
- Attach frames or photos of the real subject. It was the only thing that got the right kart.
- Say what not to write. "Don't repeat any words from my title" saves a round trip.
- Fix the shape. Crop to 16:9 at 1280×720, then check nothing important got cut.
- Read every word. "KICK $$" is a small slip that a viewer will notice.
- Check it against the real competition. Preview it in a YouTube feed beside the videos it'll sit next to, and run it through a critique for what to fix.
Key takeaways
- ChatGPT can make a strong thumbnail. Its first try scored 96 in our critique.
- Without your footage, it invents. The title-only try drew a kart that isn't in the video, and no score caught it.
- The shape is wrong by default. 1536×1024 gets black bars on YouTube.
- It repeats the title unless you tell it not to.
- Attach real frames. That fixed the kart; nothing else did.
Tested on 21 October 2026 through OpenAI's API: a chat model (gpt-6-luna) with the image generation tool set to chatgpt-image-latest, one try per message, the first image kept. Scores are from ThumbnailPeak's critique with the video's title, on the images as YouTube shows them. ThumbnailPeak is not affiliated with OpenAI, and ChatGPT is OpenAI's trademark.
The test uses "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212" by DIY PRO, published under the Creative Commons Attribution licence (CC BY 3.0). The frames are from that video. The thumbnails from Generate are adaptations made from its frames; the ones from ChatGPT's image model were made from the title, the description and four frames. All are shown unedited, and DIY PRO did not make or endorse them.
