Create and refine YouTube thumbnails without leaving the conversation. Give your AI assistant a video title plus 1–3 reference thumbnails whose style you want to match, get a finished thumbnail back in chat, then keep editing it in plain language — "make the text yellow", "swap the background for a city at night".
These are the same generations and edits as the Thumbnail Lab in your NexLev dashboard — same account, same allowance, and everything you make in chat shows up in your Lab history.
Prerequisites
- Connected AI assistant — Claude, ChatGPT, or any other supported platform
- Signed in with your NexLev account (OAuth) — Connect to your AI assistant
- Lite, Pro, or Premium plan — Free plans do not have access. Upgrade your plan to get started.
How It Works
Generate
Give your assistant the video title and 1–3 reference thumbnails whose style you like. It can pull references automatically from NexLev's research tools — for example the thumbnails of outlier videos in your niche via Trending Videos or YouTube Tools.
Optionally add creative direction ("use a red arrow", "dark background"), a short script summary for extra context, or ask for a text-free image. You can also name the format (
16:9YouTube,9:16Shorts,1:1,4:5), the resolution (1K,2K,4K) and the model — see Output Shape and Resolution.
Tools
generate_thumbnail
Generates a brand-new thumbnail from your video title plus 1–3 reference thumbnails whose style should be matched. Returns a jobId immediately — your assistant then polls get_thumbnail_generation_status until the image is ready.
Best for:
- "Make me a thumbnail for this video title"
- "Match the style of the top videos in my niche"
- "Give me a textless thumbnail I can add my own copy to"
- "Try the same title in a darker, higher-contrast style"
Example prompts:
Generate a thumbnail for my video "I Tried Waking Up at 4AM for 30 Days" — match the style of the top productivity channels
Find the 3 highest-outlier videos in the personal finance niche and use their thumbnails as style references for my video "I Saved $10,000 in 90 Days"
Generate a thumbnail for "The Truth About AI Coding Tools" using a dark background and a red arrow, based on this reference: https://i.ytimg.com/vi/dQw4w9WgXcQ/maxresdefault.jpg
Generate a no-text thumbnail for my video "5 Hidden Gems in Japan" so I can add my own headline later
Use the Pro Model to generate a thumbnail for "My $50,000 Home Studio Tour"
Generate a 9:16 Shorts cover in 4K for "This One Habit Changed My Mornings"
Use GPT Image 2.5 Flare Low to make me four cheap 1K concepts for "I Read 52 Books in a Year"
Parameters:
| Parameter | Type | Description |
|---|---|---|
newVideoTitle | string | Required. Title of the video the thumbnail is for |
referenceThumbnailURLs | string[] | Required. 1–3 HTTPS URLs of reference thumbnails (JPEG/PNG/WEBP/GIF, max 2 MB each) |
referenceThumbnailTitles | string[] | Video titles of the references, in the same order — optional, but improves results |
generationInstructions | string | Creative direction, e.g. "use a red arrow", "dark background" (max 2000 chars) |
newVideoScript | string | Script or summary of the video for extra context |
generationMode | enum | classic (full AI thumbnail with text, default) or no_text (textless image) |
imageModel | enum | Which model generates the image — see Image Models. Defaults to the best model your plan includes |
aspectRatio | enum | 16:9 YouTube (default), 9:16 Shorts/Reels/TikTok, 1:1 square, or 4:5 portrait |
outputResolution | enum | 1K (default), 2K, or 4K. The Lite Model is 1K-only; on pay-as-you-go actions the higher tiers cost more |
get_thumbnail_generation_status
Checks a generation job and returns the finished image. Your assistant polls this every 5 seconds until the job reports completed or failed.
What it returns:
status—pending,processing,completed, orfailedimageUrl— the finished thumbnail, displayed inline in the chatgenerationId— pass this toedit_thumbnailto refine the resultimageModel— which model actually produced the image (fallback-aware, so it can differ from what you asked for)aspectRatioandoutputResolution— the shape and tier the image came out at
Example prompts:
Parameters:
| Parameter | Type | Description |
|---|---|---|
jobId | string | Required. The jobId from generate_thumbnail |
edit_thumbnail
Applies an AI edit to a thumbnail — change text, colors, background, or objects — from a plain-language instruction. Returns a jobId; your assistant then polls get_thumbnail_edit_status.
Pass the edit prompt plus exactly one image source:
| Source | When to use it |
|---|---|
documentId | Continue refining the same Lab document — preferred for follow-up edits |
generationId | Edit a completed generate_thumbnail result (a Lab document is opened for it) |
imageUrl | Edit any image by HTTPS URL (JPEG/PNG/WEBP/GIF, max 10 MB — imported into the Lab) |
Example prompts:
Replace the background with a city at night and keep the person as-is
Edit this thumbnail so the headline reads "90 DAYS LATER": https://i.ytimg.com/vi/dQw4w9WgXcQ/maxresdefault.jpg
Show me my recent Lab generations and let's edit the second one — make the colors more saturated
Parameters:
| Parameter | Type | Description |
|---|---|---|
prompt | string | Required. Edit instruction, e.g. "make the text yellow" (max 2000 chars) |
documentId | string | Lab document to keep refining — from a previous edit |
generationId | string | A completed generation to edit — from get_thumbnail_generation_status or list_thumbnail_generations |
imageUrl | string | Arbitrary image to edit by HTTPS URL |
imageModel | enum | The Pro Model, the Lite Model, Nano Banana 2.1, or any GPT Image 2.5 model. The legacy GPT Image 2 models are generation-only and cannot edit. Defaults to the best model your plan includes |
get_thumbnail_edit_status
Checks an edit job and returns the edited image. Polled every 5 seconds until completed or failed.
What it returns:
status—pending,processing,completed, orfailedimageUrl— the edited thumbnail, displayed inline in the chatdocumentId— pass it back toedit_thumbnailto keep refining the same thumbnail
Example prompt:
Parameters:
| Parameter | Type | Description |
|---|---|---|
jobId | string | Required. The jobId from edit_thumbnail |
list_thumbnail_generations
Lists your past Thumbnail Lab generations, newest first, with their status, image URL, and generationId — so you can revisit an earlier thumbnail or pick one up for editing.
Best for:
- "Show me the thumbnails I made this week"
- "Find the one I generated for my morning routine video and edit it"
Example prompts:
List my recent generations and edit the newest one — make the text yellow
Parameters:
| Parameter | Type | Description |
|---|---|---|
limit | number | Results per page — 1–50, default 10 |
page | number | Page number, default 1 (newest first) |
Image Models
Six models are in service, plus two legacy GPT Image 2 models that are being retired.
| Model | Best for | Included in | AI edits | Output sizes |
|---|---|---|---|---|
| Pro Model | Highest quality generation and editing | Pro | ✅ | 1K, 2K, 4K |
| Lite Model | Fast, lightweight generation | All plans | ✅ | 1K |
| Nano Banana 2.1 | Sharper visuals and text rendering | Pro | ✅ | 1K, 2K, 4K |
| GPT Image 2.5 Flare | Fast model — all aspect ratios, generation and edits | Pro | ✅ | 1K, 2K, 4K |
| GPT Image 2.5 Flare Low | Budget tier — fastest and cheapest, lighter detail | All plans | ✅ | 1K, 2K, 4K |
| GPT Image 2.5 Sunburst | Precision model — sharpest detail, most faithful edits | Pro | ✅ | 1K, 2K, 4K |
| GPT Image 2 (legacy) | Superseded by GPT Image 2.5 Flare | Pro | — | 1K, 2K, 4K |
| GPT Image 2 Low (legacy) | Superseded by GPT Image 2.5 Flare Low | All plans | — | 1K, 2K, 4K |
Models marked Included in: Pro are still selectable on Lite and Premium — they are just billed from pay-as-you-go credits instead of your included allowance, with no effect on that allowance. See Beyond Your Allowance.
If you don't name a model you automatically get the best one your plan includes — the Pro Model on Pro, the Lite Model on Lite and Premium.
Output Shape and Resolution
Thumbnails are not 16:9-only. Every model supports every aspect ratio, so the same tools cover Shorts covers and Instagram posts as well as YouTube.
aspectRatio | Label | What it's for |
|---|---|---|
16:9 | YouTube | Standard video thumbnail — default |
9:16 | Shorts · Reels · TikTok | Vertical cover for short-form video |
1:1 | Square | Instagram feed and profile grids |
4:5 | Portrait | Instagram portrait posts |
Resolution is a separate choice: 1K (the default), 2K and 4K. Every model supports every aspect ratio, but not every model supports every resolution — the Lite Model is 1K-only. The aspect ratio never changes the price; only the resolution tier does, and only on pay-as-you-go actions.
Pixel sizes. For the GPT Image models:
| Aspect ratio | 1K | 2K | 4K |
|---|---|---|---|
16:9 | 1280 × 720 | 2560 × 1440 | 3840 × 2160 |
9:16 | 720 × 1280 | 1440 × 2560 | 2160 × 3840 |
1:1 | 1024 × 1024 | 2048 × 2048 | 2880 × 2880 |
4:5 | 1024 × 1280 | 2048 × 2560 | 2560 × 3200 |
The Pro and Lite models render slightly different 1K sizes — 16:9 ≈ 1376 × 768, 9:16 ≈ 768 × 1376, 1:1 = 1024 × 1024, 4:5 ≈ 928 × 1152 — each within 1% of the requested shape. The Pro Model's 2K and 4K tiers double and quadruple those dimensions (16:9 → 2752 × 1536 and 5504 × 3072; 1:1 at 4K → 4096 × 4096), so at "4K" it delivers more pixels than the GPT models. The Lite Model produces 1K only.
Example Workflows
Niche-matched thumbnail from scratch
Find the 3 highest-outlier videos in the "AI tools" niche from the last 30 days, then generate a thumbnail for my video "I Replaced My Whole Workflow With AI" in their style
Your assistant finds the outlier videos, reuses their thumbnails as references, generates, polls, and shows you the result.
Iterate to a final cut
Generate a thumbnail for "The Only Budget Video You Need" based on these two references, then make the text yellow, then swap the background for a plain dark gradient
Each round is a separate edit on the same Lab document — you keep every intermediate version.
Rework an existing thumbnail
Take the thumbnail from this video and change the headline to "IT GOT WORSE": https://www.youtube.com/watch?v=dQw4w9WgXcQ
One thumbnail, every format
Generate a thumbnail for "I Built a PC for $300" in 16:9 for YouTube, then make me a 9:16 version for the Shorts teaser and a 1:1 one for Instagram
Every model supports every aspect ratio, so you can cover a whole launch from one prompt — each one counts as its own generation.
A/B variants for one title
Generate two thumbnails for "I Quit My Job After 9 Years" — one with a dark high-contrast style and one bright and colorful
Included Allowance
One shared allowance across chat and the dashboard — a generation from your AI assistant counts exactly the same as one made in the web Lab. The allowance counts actions, not models: a 4K Pro Model image and a 1K Lite Model image both spend one generation.
| Free | Lite | Pro | |
|---|---|---|---|
| Generations per day | — | 3 | 5 |
| Generations per month | — | 10 | 25 |
| AI edits per day | — | 3 | 5 |
| AI edits per month | — | 15 | 25 |
| Included models | — | Lite Model, GPT Image 2.5 Flare Low, GPT Image 2 Low | All models |
| Default model | — | Lite Model | Pro Model |
| Pay-as-you-go | — | Optional (opt in) | Optional (opt in) |
Resets are in UTC: the daily allowance at 00:00 UTC, the monthly allowance at 00:00 UTC on the 1st.
What counts. Only generations and edits that actually run count against your allowance. Failed and blocked attempts don't — and neither do actions billed to pay-as-you-go credits.
Beyond Your Allowance (pay as you go)
Pay-as-you-go is optional and off until you turn it on. With usage credits enabled you can keep working after your included allowance runs out, and you can use models your plan doesn't include.
Opt in
Enable usage credits for Thumbnail Lab under Account settings → Credits (dashboard.nexlev.io/account-settings?tab=credits).
When credits are used: either after your included daily or monthly allowance is used up, or immediately — with no effect on your allowance — when you pick a model that isn't included in your plan (for example the Pro Model on a Lite plan).
Generation prices (per image)
The price already includes the resolution tier — the 2K and 4K columns are totals, not add-ons:
| Model | 1K | 2K | 4K |
|---|---|---|---|
| Pro Model | $0.16 | $0.23 | $0.33 |
| Lite Model | $0.09 | — | — |
| Nano Banana 2.1 | $0.15 | $0.18 | $0.27 |
| GPT Image 2.5 Flare | $0.21 | $0.30 | $0.45 |
| GPT Image 2.5 Sunburst | $0.21 | $0.30 | $0.45 |
| GPT Image 2 (legacy) | $0.12 | $0.18 | $0.27 |
| GPT Image 2 Low (legacy) | $0.07 | $0.08 | $0.09 |
GPT Image 2.5 Flare Low is the cheapest tier of all, topping out at $0.16 for 4K. Its exact per-tier rates are shown in the app before you generate — your assistant reports the price for the model and tier you picked.
AI edit prices (per edit)
Edits carry no resolution surcharge — they act on the document you're already working in and always output 1K:
| Model | Price |
|---|---|
| Pro Model | $0.14 |
| Lite Model | $0.07 |
| Nano Banana 2.1 | $0.07 |
| GPT Image 2.5 Flare | $0.17 |
| GPT Image 2.5 Flare Low | $0.09 |
| GPT Image 2.5 Sunburst | $0.17 |
- You're always told the price before an over-allowance action is charged.
- Without usage credits enabled, the tools simply stop at your included allowance and tell you when it resets, and models your plan doesn't include stay locked.
Limits and Rules
- Hourly ceiling. Up to 30 generation requests and 30 AI edit requests per hour, whatever your plan and even with pay-as-you-go enabled. This is an anti-runaway guard, not your allowance.
- Reference thumbnails. Every generation needs 1–3 references, each up to 2 MB, as JPEG, PNG, WEBP or GIF. Passing the reference videos' titles too is optional, but it improves results.
- Prompt length. Creative direction on a generation and an AI edit instruction are each limited to 2,000 characters.
- Images to edit. An image you ask your assistant to edit by URL can be up to 10 MB.
- AI edits always output 1K and keep the shape of the image you are editing.
- Edit guardrail. An AI edit prompt has to describe a visual change to your thumbnail. Anything else — questions for the AI, off-topic tasks, attempts to reprogram the editor — is rejected with a hint, and doesn't use up any allowance or credits. Be as vague ("make it pop") or ambitious ("redesign this") as you like; that's fine. The same check runs on generation briefs.
Troubleshooting
"This tool requires an OAuth connection"
- Thumbnail Lab is unavailable on API-key connections. Reconnect your assistant with a NexLev account sign-in — see the Quickstart.
Reference thumbnail rejected
- References must be HTTPS URLs pointing at a JPEG, PNG, WEBP, or GIF under 2 MB. If a
maxresdefault.jpgURL 404s, that video has no max-res thumbnail — tryhqdefault.jpginstead. - You need at least 1 and at most 3 references.
Job stuck on processing
- Generations and edits normally finish in 30–90 seconds. Your assistant polls for up to 5 minutes; if it times out, ask it to check the status again with the same
jobId.
Job failed
- Failed jobs are refunded automatically — no allowance or credits are consumed. Try again, and simplify very long
generationInstructionsif it keeps failing.
A model I asked for isn't available
- The Pro Model, Nano Banana 2.1, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst are included with Pro. On Lite and Premium they need usage credits enabled — turn them on under Account settings → Credits — or omit
imageModeland you'll get the best model your plan includes.
Charged credits while I still had generations left
- You picked a model that isn't included in your plan (for example the Pro Model on a Lite plan). Those are always billed from pay-as-you-go credits and never touch your included allowance. Pick an included model to use your allowance, or upgrade to Pro.
Asked for a GPT Image 2 model but got an edit error
- The legacy GPT Image 2 and GPT Image 2 Low are generation-only. For AI edits choose the Pro Model, the Lite Model, Nano Banana 2.1, or any GPT Image 2.5 model.
Asked for 2K or 4K and got a 1K image
- The Lite Model produces 1K only, so higher tiers come back as
1Kinstead of failing — switch to any other model for the higher tiers. AI edits also always output 1K by design. CheckoutputResolutionon the finished job to see what you actually got.
My edit prompt was rejected
- AI edit prompts have to describe a visual change to the thumbnail. Questions, off-topic tasks and attempts to reprogram the editor are refused — and cost you no allowance or credits. Vague ("make it pop") and ambitious ("redesign this") are both fine.
Edited image is missing my manual layers
- Expected. AI edits act on the base image; layers you added in the web editor are preserved in the Lab but not baked into the image returned in chat.
Allowance reached
- Enable usage credits in your NexLev account to keep going past your included allowance, or wait for the reset your assistant reports. Upgrade to Pro for a larger included allowance.
FAQ
Do failed generations count against my limits?
No. Only work that actually runs counts, and nothing failed is ever charged.
Can I edit with GPT Image 2?
No — the GPT Image 2 models are generation-only, and they're being retired. For AI edits choose the Pro Model, the Lite Model, Nano Banana 2.1, or any GPT Image 2.5 model.
Why can't I generate at 2K or 4K?
The Lite Model produces 1K only. Switch to another model for the higher tiers.
Do 2K and 4K cost extra?
Not while you're inside your included allowance. On pay-as-you-go, higher tiers carry the surcharge shown in the price table.
When do my limits reset?
Daily at midnight UTC, and monthly at midnight UTC on the 1st.
Do I have to enable pay-as-you-go?
No. Without it you simply stop at your included allowance until the next reset, and nothing can ever be charged.
What's Next?
- Trending Videos — Find the outlier videos whose thumbnails make the best style references
- Similar Thumbnails — Search thumbnails visually across millions of YouTube videos
- Swipefile — Save the thumbnails and videos you want to model
- Channel Analytics — Study the channels behind the styles you're matching
- Niche Overview — Understand the audience you're designing for