Prompt length limits by model
Most video models cap how long a prompt can be, and the cap is enforced by truncating from the end rather than by returning an error. The end of a prompt is usually where its exclusions sit, so an over-long prompt comes back with the text and watermarks you asked it to leave out, and nothing reports why.
The cap is a rejection threshold, not a target. Every vendor that publishes a recommended prompt length recommends far less than its own limit allows, usually a quarter to a third of it. Kling accepts 2500 characters and suggests 40 to 60 words. PixVerse reports that control drops past 80 words. Writing to the cap produces worse output, not more detail. Use the cap to decide what has to be cut, never how much can be added.
Hosted APIs
| Model | Prompt cap | Negative prompt | Vendor's recommended length | Notes |
|---|---|---|---|---|
| Kling 3.0Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 2,500 characters | 2,500 characters, separate budget | 40 to 60 words on 2.5 Turbo Pro, 60 to 100 on 2.6 and O1, 15 to 40 for image to video | Storyboard mode caps each storyboard at 512 characters, far tighter than the headline figure and the number that matters for multi-shot work. |
| Kling 2.6Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 2,500 characters | 2,500 characters, separate budget | 60 to 100 words | Same limits as 3.0. The negative prompt has its own budget, so exclusions moved there cost nothing against the prompt. |
| Luma RayVendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 6,000 characters | no negative field | 40 to 100 words | The most generous cap of any hosted model, and the widest gap between what is allowed and what is advised. Prompts outside 1 to 6000 characters are rejected outright. |
| PixVerse V6Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 2,048 characters | Field exists, no limit published | 50 to 80 words | PixVerse does not publish a character limit for either field. Its API documentation lists prompt and negative_prompt in the request body and states no constraint on the length of either, so the 2048 figure is what aggregators enforce rather than what PixVerse specifies. Treat it as the practical ceiling, not a vendor number. PixVerse's own testing reports that control drops past 80 words. |
| Vidu Q2Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 3,000 characters | no negative field | none published | Providers disagree. Vidu's own reference says 3000 characters and some resellers document 1500. Write to the lower figure until the vendor settles it. |
| Sora 2Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 2,000 characters | no negative field | Under 1500 characters | Sources disagree on whether the unit is characters or tokens. Semantic drift is reported past roughly 1500 characters, where the model starts following late instructions and dropping early ones. |
| Hailuo 02Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 2,000 characters | no negative field | none published | The same 2000 character limit applies to text to video and image to video. |
| Runway Gen-4Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | 1,000 characters | no negative field | none published | The tightest cap among the major hosted models, which makes it the portable figure: a prompt that fits Runway fits everywhere. |
| Veo 3Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | not published | Field exists, no limit published | 100 to 150 words, or 3 to 6 sentences | Google publishes a recommended length rather than a cap. A negativePrompt field exists. No published cap is not the same as no cap, so assume one and keep headroom. |
| Veo 3.1Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | not published | Field exists, no limit published | 100 to 150 words, or 3 to 6 sentences | As Veo 3. The prompting guide advises describing subject, context, action and style, with camera and audio optional. |
| Seedance 2.0Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | not published | no negative field | none published | The published API schema states no length constraint and offers no negative prompt field, so exclusions have to sit in the prompt itself. |
If you only remember one number
1,000 characters. Runway has the tightest cap among the major hosted models. A prompt written to 1000 characters runs on any of them without a rewrite, which matters when you switch model or price-shop between providers.
Models you run yourself
Models you run yourself have no character cap. They have a text encoder token limit instead, and it fails differently: the prompt is cut at the encoder before the model ever sees it, with no error and no warning. 512 tokens is roughly 1800 to 2000 English characters.
| Model | Text encoder | Token limit | Notes |
|---|---|---|---|
| Wan 2.2Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | UMT5-XXL | 512 tokens | Truncated beyond 512 tokens. Users report quality degrading well before that, from around 320 to 350 tokens, where motion slows and a grid artefact appears. |
| HunyuanVideoVendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | T5-XXL | 512 tokens | Commonly configured at 512 tokens. Verify against the implementation you are running, because the limit is set by the pipeline rather than the weights. |
| LTX-2.5Vendor figure, checked 2026-08-16Collected from the vendor's own documentation on 16 August 2026. Vendors change these between model versions without announcing it, so check the vendor's current API reference before you rely on it. Check the vendor's documentation | T5 v1.1-XXL | not documented | No token limit documented for LTX-2. Handles prompts of 100 words and more comfortably. |
Getting more into the budget
Move exclusions into the negative field where one exists
Kling gives the negative prompt its own 2500 characters and PixVerse its own 2048, so exclusions moved there cost nothing against the prompt. Do not put them in both: it spends the budget twice, and some models read a repeated ban as emphasis on the banned thing.
State a shared decision once per prompt
In a multi-shot prompt, style, cast and audio belong in one header above the shots rather than inside each one. Six shots repeating a 300 character header spend 1800 characters saying the same thing six times. It also makes the result worse, because restating a decision invites the model to reinterpret it mid-prompt. The exception is when each shot is a separate generation, where every prompt must carry its own full copy.
Image to video wants the shortest prompt of all
15 to 40 words on Kling. The frame already carries the subject, the setting and the colour, so the prompt only has to carry motion. Redescribing what is already in the frame invites the model to redraw it.
Count, do not estimate
It looks about right is how a prompt arrives 300 characters over and comes back missing its exclusions. Count the characters before you send it.
Common questions
What is the prompt character limit for AI video models?
It varies by model and most vendors publish it in their API reference rather than their marketing pages. Kling allows 2500 characters, Sora 2 and Hailuo 02 around 2000, PixVerse 2048, Vidu Q2 3000, Luma Ray 6000, and Runway 1000. Veo and Seedance publish no cap. Write to 1000 characters if you want one prompt that runs on all of them.
What happens if a prompt is too long?
Most models truncate it from the end rather than returning an error. The end of a prompt is usually where its exclusions sit, so the clip comes back containing the text, watermarks or extra limbs you asked it to leave out, and nothing tells you the prompt was cut.
Should I write prompts up to the character limit?
No. Every vendor that publishes a recommended length recommends a quarter to a third of its own cap. Kling allows 2500 characters and suggests 40 to 60 words. PixVerse reports control dropping past 80 words. The cap is where the request is rejected, not where quality peaks.
Does the negative prompt count against the limit?
Not where the model gives it a separate field. Kling allows 2500 characters of negative prompt on top of the 2500 for the prompt, and PixVerse 2048 on top of 2048. Where a model has no negative field, such as Seedance, exclusions come out of the prompt budget and belong at the end.
Do open weights models have a prompt limit?
They have a text encoder token limit rather than a character cap. Wan 2.2 truncates at 512 tokens through its UMT5-XXL encoder, roughly 1800 to 2000 English characters, and users report quality degrading from around 320 tokens. The prompt is cut before the model sees it, with no warning.