Skip to main content

AI Video Generation Guide: Runway, Sora and Veo Today

AI video generation guide: where Runway, Sora and Veo stand now, how the four generation modes work, prompt rules from the vendors, and what one clip costs.

Founder of IImagined.ai

Published
Apr 5, 2026
Updated
Oct 13, 2026
Reading time
10 min read
Quick answer

Of the three tools in this guide's name, two are current. Runway's model for new clips is Gen-4.5: 2 to 10 seconds at 720p, silent, on the $15 Standard plan. Google's Veo 3.1 makes 4, 6 or 8 second clips with sound, with a free route through Google Flow at 50 credits a day. Sora is gone: OpenAI closed the app on April 26, 2026 and the API on September 24. The method is the same on every tool: settle the look in a still, describe one action per clip, and join shots in an editor. Figures were read on the vendors' pages on 13 October 2026; we did not test output quality.

AI video generation today means short clips, 4 to 10 seconds at a time, made from a text prompt or a still image and joined in an editor. Of the three tools this guide is named for, Runway (Gen-4.5) and Google Veo (3.1) are current, and Sora has been shut down. The steps below work on either, and on the newer models that have arrived since.

What this page is based on: Runway's Gen-4.5 article, prompting guide and pricing; Google's Veo docs, Flow credits page and API pricing; and OpenAI's Sora discontinuation article. All were read on 13 October 2026. Prices are US dollars. We did not generate clips, so nothing here is a quality score.

Runway, Sora and Veo: where each stands now

Two of the three still exist, and both have moved on from the versions older guides describe. Runway Gen-3 and Veo 2 are not what you would open today.

ToolCurrent modelOne generationSoundHow you pay
RunwayGen-4.5 for new clips. Aleph 2.0 for editing existing video2 to 10 seconds at 720p, 24 or 25 fps. Text to video or image to videoNone listed for Gen-4.5Standard plan and higher: $15 a month for 625 credits. 12 credits a second
SoraNone. The app closed on April 26, 2026 and the API on September 24, 2026Not availableNot availableNot for sale. OpenAI lists no replacement
VeoVeo 3.1, in Lite, Fast and Quality tiers. Google now points new work to Gemini Omni Flash first4, 6 or 8 seconds at 720p, 1080p or 4K (no 4K on Lite), 24 fpsGenerated with the videoGoogle Flow: 50 free credits a day. Lite 10 credits a clip, Fast 20, Quality 100

Sources: Runway's Gen-4.5 spec table and pricing page, OpenAI's deprecations page, and Google's Veo docs and Flow credits page.

Where to go from Sora depends on what you used it for; our Sora alternatives page maps each job to a replacement. Two more things changed on the Google side. Its video docs now say to use Gemini Omni Flash "as your default model for video generation" and to keep Veo 3.1 for scene extension, last-frame control and older pipelines. And Runway is no longer one model: its API price list includes other vendors' models, Veo 3.1 among them, and it describes Aleph 2.0 as its in-context video editing model.

How AI video generation works: four modes

Every tool offers some mix of the same four modes, and choosing the mode matters more than choosing the brand. You either start from words or from a picture, and you either make a new clip or change one you have.

The four modes of AI video generation
Change an existing clip
Text to video. The model invents the look and the motion. Fastest to try, hardest to repeat.
Image to video. A still fixes the look and the prompt describes only the motion. The mode to use for a recurring character or product.
Make a new clip
Prompt editing. A follow-up instruction changes a clip: Gemini Omni Flash by a new turn, Runway with Aleph 2.0.
Extension and frame control. Continue a clip, or set a first and last frame and let the model fill the middle. Veo 3.1 does both.
Start from text
Start from an image or clip

For anything you plan to publish more than once, work in the right-hand column. A still you approved holds the face, the product and the lighting, so each attempt only risks the motion. That method is laid out in our image-to-video guide.

How to write a prompt for AI video

Start with three things, a subject, an action and a style, and add the rest only when the shot needs it. That is the structure in Google's Veo prompt guide, which lists seven elements. The examples in the last column are ours, to show the idea.

ElementWhat Google says it isIllustrative example
SubjectThe object, person, animal or scenery you want in the videoA ceramic mug on a wooden counter
ActionWhat the subject is doingSteam rises as coffee is poured in
StyleCreative direction, using film style keywordsWarm documentary look
Camera position and motion (optional)Where the camera is and how it movesSlow push in at counter height
Composition (optional)How the shot is framedClose-up, mug on the left third
Focus and lens effects (optional)Specific visual effectsShallow depth of field
Ambiance (optional)How colour and light contributeSoft morning window light

Veo also takes sound direction in the same prompt. Google's guide says to put specific speech in quotes, to describe sound effects explicitly, and to describe the soundscape of the place. Its two writing tips are plain: use descriptive language, and include the word "portrait" when facial detail matters.

When you start from an image, describe the motion

Runway's prompting guide makes the point that the image is already part of the prompt. Repeating what is in the picture can reduce motion. It recommends beginning with the simplest prompt that captures the essential motion, then adding subject motion, camera motion, scene motion and style one at a time, so you can tell which addition helped.

Runway's prompting rules, from its own guide
Avoid
  • Negative phrasing: "No camera movement"
  • Abstract language about a feeling or an idea
  • Re-describing the hair, suit and tie already in the image
  • Requests and commands: "can you please add my dog"
  • Several scene changes and style shifts in one clip
Do
  • Positive phrasing: "Locked camera. The camera remains still."
  • A physical action: "The woman smiles and waves."
  • Motion only: "The man extends his arm to shake hands, then nods."
  • A description: "A dog excitedly runs into the scene from off-camera"
  • One scene per generation, in general terms like "the subject"

A workflow from prompt to published video

The order that wastes the fewest credits is still first, motion second, sound and edit last. Each step below changes one thing, so a bad result has one cause.

Seven steps from idea to upload
  1. 1
    Write the shot list

    One line per clip: subject, action, camera. A 30-second video is four to six clips, not one prompt.

  2. 2
    Make the first frame as a still

    Settle the look in an image tool, where a retry is cheap. For a person, use a fictional adult persona or someone with written permission.

  3. 3
    Animate with one action per clip

    Describe the motion only. One subject, one action, one camera move.

  4. 4
    Draft cheap, then render once

    In Flow, a 360p Omni Flash draft costs about half of 720p. Test framing there before you spend on the keeper.

  5. 5
    Change one thing per retry

    Keep a log of prompt and settings. Runway advises adding one element at a time for the same reason.

  6. 6
    Edit, add sound, export

    Join the clips, add voice or music where the model gave none, and export in the aspect ratio of the platform.

  7. 7
    Label and archive

    Turn on the AI disclosure where the platform asks for it, and save the files and the prompt log outside the tool.

Step two is where a recurring character is won or lost. A LoRA or a set of reference images keeps the face stable across stills, and AI video character consistency covers keeping it stable across clips. For voice, see AI voice for video.

What one AI video clip costs

Per-second pricing makes the tools easy to compare, and the spread is wide: the same eight seconds costs eight times more on the top Veo tier than on the lightest. These are API list prices, which is the cleanest like-for-like figure.

One 8-second clip at 720p, by API price
Veo 3.1 Lite, with sound
$0.40
Veo 3.1 Fast, with sound
$0.80
Runway Gen-4.5, silent
$0.96
Veo 3.1, with sound
$3.20

Source: Calculated: 8 seconds times the per-second rate on the Gemini API pricing page and the Runway API pricing page (12 credits a second at $0.01 a credit), read 13 October 2026. Failed attempts are not counted.

Inside the apps you pay in credits. In Google Flow, every account gets 50 credits a day, a Veo 3.1 Lite clip costs 10, Fast 20 and Quality 100, and Google AI Plus at $4.99 a month in the United States adds 200 credits. On Runway, the Free plan is 125 credits once and Gen-4.5 needs Standard: $15 a month, or $12 a month billed yearly, for 625 credits, which is about 52 seconds of Gen-4.5.

Budget for retries. If you keep one take in three, the real price of a finished clip is three times the figure above. Free allowances across more tools are in our free AI video generator comparison, and paid plans side by side are in the AI video generator comparison.

How to make a video longer than one clip

You extend, or you join. Two models extend their own output; everything else is editing.

MethodHow it worksLimit
Veo 3.1 extensionAdds 7 seconds to a Veo clip, up to 20 timesUp to 148 seconds. 720p only, and not on Veo 3.1 Lite
Gemini Omni 1.1 Flash extensionAdds 10 seconds at a timeA total of 40 seconds
Join shots in an editorAny model, any lengthYou handle continuity: reuse the last frame as the next first frame

Source: Google's Veo docs and Omni Flash docs. One practical warning from the Veo docs: generated videos are stored on Google's server for 2 days and then removed, so download each result when it finishes.

Rights, watermarks and labels

Publishing is where the rules sit. Check four things before a generated clip goes out.

Before you publish an AI-generated video
  • Likeness: the person on screen is you, a fictional adult persona you created, or someone who gave written permission.
  • Plan terms: read what your plan says about commercial use. Google says it won't claim ownership of what you generate in Flow.
  • Watermark: Google says videos created by Veo carry an invisible SynthID watermark. Do not crop or blur a visible one.
  • YouTube: disclose realistic content that is generated or meaningfully altered with AI, using the AI use setting in YouTube Studio.
  • No disclosure needed on YouTube for clearly unrealistic scenes, or for AI help with scripts, titles and thumbnails.
  • Files: keep the prompt, the input image and the export, in case a platform or client asks how it was made.

Sources: the Flow FAQ, the Veo docs and YouTube's disclosure page. Instagram's version of the rule is in our Instagram AI label guide.

The self-hosted route

If per-clip pricing does not suit you, open models run on your own graphics card with no credits. Wan 2.2 is released under Apache 2.0 with text-to-video and image-to-video models at 480p and 720p, and it runs in ComfyUI. The cost moves to hardware: the repository says its smaller TI2V-5B command needs a GPU with at least 24GB of VRAM.

To set that up, install nodes with the ComfyUI Manager guide, then follow the Wan 2.2 Animate tutorial. Runway's own references feature is covered in Runway Gen 4 for creators.

Where to go next

Pick one tool, make five clips from one still, and keep the log. That teaches more than reading another comparison. When you are ready to build a persona around the clips, AI Influencers, one of our programs, has lessons on AI video generation with Kling and Runway and on image-to-video pipelines, and the AI influencers library lists every related guide.

What changes often

Model names and credit costs. Since this guide first went up in April, Sora's app and API have both closed, and Google's docs now put Gemini Omni Flash ahead of Veo as the default. The method above did not change. Check the linked vendor page before you rely on a price.

AI video generation: FAQ

Which is better for AI video generation, Runway or Veo?

They suit different jobs. Veo 3.1 generates sound with the picture, reaches 1080p and 4K, and has a free route through Google Flow at 50 credits a day. Runway Gen-4.5 is silent, outputs 720p clips of 2 to 10 seconds and needs the Standard plan at $15 a month, but Runway plans also include other vendors' models, such as Kling 3.0. Figures read 13 October 2026.

Is Sora still available?

No. OpenAI's help center says the Sora web and app experiences were discontinued on April 26, 2026, and its deprecations page lists the Videos API and the Sora 2 models as shut down on September 24, 2026, with no recommended replacement. Content made in the app can be exported at sora.chatgpt.com/sunset until OpenAI closes that window.

How much does AI video generation cost?

It can start free: Google Flow gives 50 credits a day, and a Veo 3.1 Lite clip costs 10. Google AI Plus is $4.99 a month in the United States and Runway Standard is $15 a month for 625 credits, about 52 seconds of Gen-4.5. Through the Gemini API, Veo 3.1 Lite bills $0.05 a second at 720p. Read 13 October 2026.

How long can an AI-generated video be?

One generation is short. Veo 3.1 makes 4, 6 or 8 seconds, Gemini Omni Flash up to 10, and Runway Gen-4.5 from 2 to 10. Longer pieces come from extension or editing: Veo 3.1 can extend its own 720p clip by 7 seconds up to 20 times, for as much as 148 seconds, and anything else is joined in an editor.

How do I write a good AI video prompt?

Name a subject, an action and a style, then add camera, framing and light only if they matter. That is the structure in Google's Veo prompt guide. When you start from an image, Runway's guide says to describe the motion and not the picture, to use positive phrasing, and to begin simple and add one element at a time.

Do I have to label AI-generated video?

On YouTube, yes, when it looks real. YouTube requires creators to disclose realistic content that is generated or meaningfully altered with AI, through the AI use setting in YouTube Studio. Clearly unrealistic scenes and production help such as scripts or thumbnails do not need it. Creators who consistently fail to disclose can have a label applied or face penalties.

Free · no card needed

Start with the free Starter Kit.

An AI character bible from AI Influencers, plus four more working files.

Starter Kit5 files · free
See All Access: $99/mo →30-day money-back guarantee