How to make an AI video

What it costs per second

Prices vary by a factor of ten or more. Here is what a handful of popular models cost per second at around 720p, as each vendor states it.

Vendor-stated rates; hosts can add minimums or charge for reference inputs. More rows and hosts on the price tracker and in the cost calculator.

Step 1: Plan in shots, not in videos

Every current model works in short bursts. One Seedance 2.5 guide sums up the trade-offs as Kling 3.0 for multi-shot scenes of 3 to 15 seconds and Veo 3.1 for 8-second clips; each model page lists its own limits. So the real first step happens on paper. Write the piece as a list of shots, one line each: "wide: a cyclist crests a hill at dawn", "close-up: her breath in the cold air", "drone: the road unspooling below".

Every prompt guide we cite below assumes you work this way. A model given "a short film about a cyclist" has to invent the story, the cuts and the camera. A model given one shot only has to do one thing well.

Step 2: Pick a model and a place to run it

There are two ways in. Apps (Dreamina, CapCut, Kling's site, platforms that bundle several models) give you a text box, buttons and usually some free credits. APIs (fal, Runway's API, Google's Gemini API, xAI and others) charge per second of video and suit people who generate a lot or want to automate. Start in an app; switch to per-second billing once you know which model you like and are making enough to care about the bill.

Prices vary by a factor of ten or more. Here is what a handful of popular models cost per second at around 720p, as each vendor states it:

Price is not quality, and we don't rank models here. A sensible approach is to draft on something cheap and re-run the shots that matter on a stronger model. If you're choosing between two, the side-by-side compare pages list modes and prices together.

Step 3: Write the prompt like a shot description

The guides we've read agree more than they disagree. Describe the subject, then the action, then the camera, then the setting and light, then the style, then the sound. A Seedance 2.5 guide suggests global rules for camera position, lighting, colour and ambience, and fal's Seedance 2.5 template builds each stage around one primary action or event. A worked example:

Too vague: a woman cycling, cinematic, 4K, beautiful

Better: A woman in a red rain jacket stands on her pedals as she crests a hill at dawn. Low camera at road level, slow push-in as she passes. Mist in the valley behind her, warm low sun from the left. Natural colour, light film grain. Sound: wind, chain clicking, distant birds.

Two habits save the most money. First, set length, aspect ratio and resolution in the settings, not in the text; one Seedance 2.5 guide calls spec tokens like “4K, 60fps” “noise; choose resolution in the platform.” Second, describe what you want rather than listing what you fear. The same guide's advice on negative prompts: “write a CONSTRAINTS line instead.”

More examples by model: Seedance prompts, Kling prompts and Veo prompts.

Step 4: Start from a picture when the look matters

Text-to-video is fine for exploring. When you need a specific face, product or style to stay put, generate or shoot a still first and animate it (image-to-video). Then write only what changes. fal's Seedance 2.5 guide says the same for reference video: when the reference already defines the motion, “state only which attributes to inherit.” The image or clip sets the look; spend the words on movement, expression, weather and camera.

Step 5: Generate a few takes, change one thing

Plan on two or three takes per shot. When a take is close but wrong, resist rewriting the whole prompt. Change a few words at a time so you can tell which change helped. If the camera is wrong, change the camera sentence; if the mood is wrong, change the light or style sentence.

Some failures are well known and have known fixes: the camera ignoring your move, faces drifting between shots, products changing shape, unwanted cuts, burned-in text. We keep sourced fixes for these by model on Fix a failed take, for example face drift in Seedance and Kling ignoring the camera.

Step 6: Edit it like normal footage

Bring the keepers into any editor (CapCut, Premiere, DaVinci Resolve, even a phone app), trim each clip to its best moment, and cut on action. If your model generated sound, check it shot by shot; a music bed or ambience track across the whole edit can smooth over differences between clips.

What it costs: a worked example

Say you're making a 30-second piece: six shots of 5 seconds, three takes each. That's 90 seconds of generated video. At the rates above:

• Veo 3.1 Lite (720p, with audio): Discontinued: Google listed these preview IDs as shutting down Oct 22, 2026; use gemini-omni-1.1-flash (Gemini Omni Flash) on the Gemini API, or the GA Veo 3.1 models on Gemini Enterprise Agent Platform (veo-3.1-generate-001, veo-3.1-fast-generate-001).
• Kling 3.0 (720p, audio): 90 × $0.126 = $11.34
• Seedance 2.5 via Runway (720p output): 90 × $0.30 = $27.00, plus Runway's charge for any input video
• Seedance 2.5 on fal (720p): 90 × $0.473 = about $42.57 (fal states the rate as “roughly”)

The takes, not the final length, drive the bill, which is why tighter prompts are the best saving. Run your own numbers in the calculator.

A simple first project

If you want somewhere to start tonight: pick one subject you can describe in detail, write four shots (wide, medium, close-up, a moving shot), generate each twice on a cheap model, keep the best four, and cut them to a 20-second piece of music. You'll learn more from that than from any single perfect clip. When you're ready for longer work, our workflow guides cover multi-shot pipelines.

Prices are vendor-stated and dated in the table; quotes are checked against the saved source pages. No links on this page are paid.

When you need a pick by job (not a tutorial), open best AI video generators compared.

FAQs