How to use Veo 3.1.
Every mode, every limit, every credit.
Veo 3.1 needs a Playcut account on a paid plan and a credit balance. No Google Cloud project, no provider key, no per-model unlock — but video is gated to paid tiers, so an unsubscribed workspace is capped to image tools. It runs from the studio, the REST API and MCP at 160 credits a second at 720p and 1080p.
16:9 and 9:16 · 4 / 6 / 8 seconds · 720p, 1080p, 4K · MP4 with generated audio · rates in credits
What you need to run Veo 3.1
A paid plan, credits, and one of three surfaces. That is the whole list. There is no per-model unlock on our side and no separate Google key to paste in, but video generation is a paid-tier tool: a workspace with no active subscription resolves to free-tier limits, which allow the two image tools and refuse every video submission with a 403.
Past that gate nothing about the model is withheld by tier. Hobby carries the same Veo 3.1, the same five modes and the same API access as Agency; what a plan changes is how far your credits go and how many renders run at once.
In the studio
Open app.playcut.ai, start a session, and generate. Veo 3.1 is the pre-selected video model; if you want another one you open the picker and choose it. The pick is manual and stays manual — nothing reroutes your job to a different model or a different rate.
Over the REST API
POST /api/v1/tools/{name}/execute with an Authorization: Bearer pk_live_… header. Keys are created in Settings → API Keys, shown once, and capped at 20 active per workspace.
Over MCP
The same five tool names are exposed to Claude, Cursor and ChatGPT through the hosted Playcut MCP server — 34 public tools in total. Setup lives on the MCP and API page.
The cheapest real request, not a spec-sheet one. Omit model and you get Veo 3.1; omit workerId and a session is created for you; omit resolution and ratio and you get 720p at 16:9. Field tables and the per-mode call set live in the Veo 3.1 developer reference.
# The smallest sum you can spend on Veo: 4 seconds at the default 720p.
curl -X POST https://api.playcut.ai/api/v1/tools/generate-video-from-text/execute \
-H "Authorization: Bearer $PLAYCUT_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt":"Espresso pouring into a glass over ice, overhead, steam rising","durationSeconds":4}'
# 200 ack — the hold has already happened, the video has not.
# {"taskId":"65f3a1b2c3d4e5f600000001","status":"PENDING","estimatedCredits":640}
The call returns a task id, not a video, and estimatedCredits tells you what was just held. Poll get-task-status until it reports a terminal state, then read the asset off the completed task. A failure reverses the hold, so a rejected render costs nothing.
Pick the mode before you pick the settings
Veo 3.1 is the only video model on Playcut that covers all five generation modes. Every other model here is missing at least one of them, which is the real reason to pay its rate. The mode you choose also decides which settings you are still allowed to touch.
Text to video
generate-video-from-text
Reach for it when: A shot that does not exist yet — no plate, no still, nothing to match.
The constraint that decides it: One of only two modes that still lets you pick 4, 6 or 8 seconds at 720p.
Animate a still
generate-video-from-image
Reach for it when: You already have the frame and only need it to move. The image becomes the first frame.
The constraint that decides it: Exactly one source image. It keeps the 4/6/8 choice at 720p, like text-to-video.
Reference to video
generate-video-from-reference
Reach for it when: A subject or a look has to carry across shots and you have images that define it.
The constraint that decides it: The only mode that pins the ratio too: 16:9 only, 8 seconds only, 3 reference images max.
Interpolation
generate-video-interpolation
Reach for it when: Storyboard to motion — you know where the shot starts and where it ends.
The constraint that decides it: Exactly two frames, a start and an end, and always 8 seconds.
Extension
generate-video-extension
Reach for it when: Pushing a finished Veo clip past its 8-second ceiling, 7 new seconds at a time.
The constraint that decides it: 720p and 8 seconds only, and the source has to be a clip Veo itself generated.
Not a Veo mode
Video editing — a source clip plus a prompt that transforms it — does not exist on Veo 3.1 at all. That runs on Wan 2.7 or HappyHorse 1.1. Audio-driven lip sync, where you supply the voice track, is Wan 2.7 only and is rejected on every other model.
What Veo 3.1 costs
160 credits per second at 720p and 1080p, 240 at 4K. Your balance is debited in credits, and every figure on this page is the credit figure — what we charge, not a list price from anyone else.
Billing is duration × rate, held at submission. Resolution changes the rate only at 4K; 1080p costs exactly what 720p costs and simply forces the longest duration.
What a plan's credits buy in Veo seconds
This is the arithmetic that makes the top-up packs make sense. A monthly allowance divided by 160 credits is a small number of seconds, and on the entry tier it does not reach a single finished clip.
Credit packs never expire and stack on top of the monthly allowance, in 600 · 2,500 · 5,000 credit sizes. Full tier breakdown on the pricing page, and the credit model in plain language in the Playcut pricing breakdown.
The rules Veo enforces before you spend
Every combination below is refused before any credits are held, so a wrong setting costs you nothing but a round trip. A bad setting comes back as a 400 and a blocked prompt as a 422; either way the balance is untouched. That ordering is deliberate — the alternative is holding credits, dispatching, and failing upstream.
- Any aspect ratio that is not 16:9 or 9:16, on text-to-video and animate. Square, 4:5 and 21:9 are refused at the door.
- Any duration that is not 4, 6 or 8 seconds. There is no 5-second Veo clip and no 12-second one.
- 4 or 6 seconds at 1080p or 4K. Both premium tiers require exactly 8 seconds, so there is no short 4K render to test a look with.
- More than 3 reference images, or reference mode at anything but 16:9 and 8 seconds. Reference is the strict one: it refuses 9:16 as well, so 16:9 is the only ratio it takes. The other modes still reject everything outside 16:9 and 9:16.
- Extension at anything but 720p and 8 seconds, on a source longer than 141 seconds, or on a clip already extended 20 times.
- A personGeneration value the mode does not accept. Text-to-video takes only allow_all and animate takes only allow_adult, so the safe move is to omit the field.
- A prompt our safety filter blocks. That one comes back as a 422 rather than a 400, on every surface. The separate 8,000-character prompt cap is a tools-API and MCP schema limit only — the studio's own REST path does not length-check your prompt.
Honest limits: what Veo 3.1 will not do
These are the limits that move a budget or a schedule, not the full parameter list. Each one is enforced by Playcut code or by one of our adapters, and each one has cost somebody a render. The spec-sheet version lives in the developer reference; this is the version that decides whether Veo fits the brief.
Limits you hit on your first clip
- A second ratio is a second render, not a re-export
- 16:9 and 9:16 are the whole list — no 1:1, no 4:5, no 21:9. A feed cut is therefore a fresh generation on a model that takes the ratio, or a crop with the loss that implies. HappyHorse 1.1 is the only model here that accepts 4:5, at 56 credits a second, so an 8-second 4:5 companion cut runs 448 credits on top of the Veo master. Price both before you commission either.
- The duration enum is the whole cost decision
- At 720p you get three rungs and three prices: 640, 960 or 1,280 credits. There is no five-second clip and no twelve-second one, so a longer piece is a sequence you assemble rather than a number you raise. Pick the rung deliberately — going from 4 to 8 seconds doubles the bill on the most expensive video model we sell.
- There is no cheap way to look at 4K
- 1080p, 4K, reference, interpolation and extension all force the full 8 seconds, which means the premium tiers have no test render. One 4K attempt costs about 3.8 months of Hobby's entire credit allowance. The workable order is to settle the composition at 4 seconds and 720p, then spend the 4K pass once, on a shot you already like.
- Three reference images is the floor of our catalog
- Veo takes three — the fewest of any Playcut video model that accepts references at all. Seedance 2.0 and HappyHorse 1.1 take nine, Grok Imagine seven, Gemini Omni Flash six, Wan 2.7 five. If the brief is a character sheet plus a product plus a palette, the slots run out before quality is even the question.
- You pay for audio whether the brief wants it or not
- Veo scores every clip in the same pass as the picture and Playcut exposes no toggle, so a clean plate for your own composer means muting in the edit. Native audio is also a weak reason to choose Veo: Seedance 2.0, Gemini Omni Flash, HappyHorse 1.1 and Wan 2.7 all generate it, at a fraction of the per-second rate.
- A rerun is a new bill, never a reproduction
- You may send a seed and we record it on the asset for provenance, but it is never forwarded to Google. So the near-miss you wanted to nudge cannot be recovered — you pay full rate for a different clip. On a model at 160 credits a second that is the limit most likely to cost you money, so lock the look with reference images or an interpolation start frame instead.
Limits that only surface later
- Extension only continues Veo output — and you find out late
- An uploaded MP4 cannot be extended, and neither can another model's output. Unlike the other rules on this page, that one is not checked at submission: it lives in our worker at the provider call, so the request is accepted, credits are held, the job dispatches and then fails. The hold is reversed in full, so you lose the wait rather than the credits — but confirm the source came out of Veo before you queue it.
- Extension is 720p only, and seven seconds bill as eight
- A 1080p or 4K clip is a dead end — there is no path to continue it. Each call adds 7 new seconds of footage but bills the 8-second output at 1,280 credits, so continued runtime effectively costs about 183 credits a second rather than 160. Two ceilings are checked before the hold: sources over 141 seconds, and clips already extended 20 times.
- personGeneration accepts exactly one value per mode
- Text-to-video takes allow_all; animate takes allow_adult. Anything else, including dont_allow, is a 400. The field is best omitted entirely, which is what the studio does.
- No video editing, and no lip sync to audio you supply
- Veo cannot take a source video plus a prompt and transform it — that is Wan 2.7 and HappyHorse 1.1. The driving-audio path, where you upload the voice track and the mouth follows it, is Wan 2.7 only and rejected everywhere else.
- It is the most expensive second of video we sell
- At every resolution tier, no other Playcut video model costs more per second. A Hobby month's 500 credits does not cover one 4-second clip, and Hobby runs one job at a time, so iterating on an 8-second shot is serial on the entry tier.
When to reach for something else
Six other video models share the same credit balance, the same workspace and the same picker. The right-hand column is the specific thing each one does that Veo 3.1 refuses — an aspect ratio, a clip length, a reference count, an input Veo has no field for. All six are cheaper per second.
The full price list for all 13 model surfaces, including images, music and voice, lives on the AI models comparison hub.
Why the duration ladder is the first thing to plan around
Most people arrive at Veo 3.1 expecting a slider and find an enum. Four, six or eight seconds is the entire menu, and the moment you ask for 1080p, 4K, reference mode, interpolation or extension, the menu collapses to eight. That single rule reshapes how you budget a sequence.
The practical method is to storyboard in eight-second beats rather than in scenes. A 32-second spot is four generations, not one long render, and the joins are where your edit does the work. Trying to stretch a single beat past eight seconds is what pushes people into extension, where the constraints get tighter rather than looser.
The second planning input is ratio. Because Veo only produces 16:9 and 9:16, a campaign that also needs a 4:5 feed cut cannot simply re-export — the 4:5 version has to be generated on a model that supports it, or cropped in post with the loss that implies. Deciding the delivery ratios before the first render saves a re-shoot.
The cheapest way to find out whether a shot works
At 160 credits a second, a wrong idea is expensive on Veo. The workflow most teams settle into is to draft on a cheap model and finish on the flagship, because prompts port between models with light editing and everything lands in the same workspace.
Grok Imagine at 20 credits a second answers "does this composition read at all" for a fraction of the cost. Seedance 2.0 at 10 credits at 480p is cheaper still, with the caveat that it refuses input images containing human faces.
Once the draft answers the question, re-run the winning prompt on Veo 3.1 at the delivery resolution. The cheap pass is not a compromise — it is the reason the expensive pass only happens once. The broader version of this argument is on the AI video generator page.
How credits are held, charged and returned
Credits are held the moment you submit, not when the render finishes. The hold is duration times the per-second rate for your resolution, taken atomically before the job is dispatched, so parallel submissions cannot push a workspace negative.
If the job fails or is cancelled, the hold is reversed in full. If it completes, the hold was the charge and nothing further happens. And if a setting is invalid — a bad ratio, a 4-second 4K request, an oversized extension source — the request is refused before any hold exists, so it costs nothing at all.
That ordering is why this page spends so long on the refusal rules. On a model this expensive, knowing which combinations are impossible is worth more than knowing which ones are optimal.
What is ours to verify, and what is Google's
Everything above is enforced by Playcut code we can point at: our validators, our adapters, our pricing table. That is the standard we hold this page to, and it is why some commonly-quoted Veo facts are missing from it.
Watermarking, the provider-side prompt token cap, frame rate, generation latency and Google's regional person-generation policy are all Google's behaviour, not ours. They can change without a deploy on our side, so we do not restate them as Playcut facts. Google's own Veo API documentation is the authority for those.
One thing worth saying plainly: there is no automatic model router on Playcut. Each capability ships a default — Veo 3.1 for video — and you change it by opening the picker. Nothing reads your prompt and quietly moves the job to a different model with a different rate.
Where this page stops and the docs begin
This page is for deciding: which mode, what it costs, whether the limits fit your brief. Field names, complete parameter tables, error semantics and the full curl set for all five modes belong in the developer reference instead of here.
Start with the Veo 3.1 model reference in the Playcut docs for the spec sheet, and the MCP server and REST API page for keys, tool names and client setup. If you are new to the studio itself, the guide to using Playcut walks the whole flow.
Frequently asked questions
Do I need a Google Cloud project to use Veo 3.1? +
No. There is no provider key to paste in and nothing to provision on Google's side. What you do need is a Playcut account on a paid plan plus a credit balance, because video generation is gated to paid tiers. Veo 3.1 is the default video model, so a subscribed workspace renders on it without changing a setting.
Can I use Veo 3.1 without a subscription if I buy credits? +
No. A workspace with no active subscription falls back to free-tier limits, and those allow the two image tools only. Every video submission is refused with a 403 before anything is held, so a credit balance on its own will not render a Veo second. Any paid tier lifts the gate.
What does one Veo 3.1 clip cost? +
160 credits per second at 720p and 1080p, 240 at 4K. So an 8-second 720p clip is 1,280 credits and an 8-second 4K clip is 1,920. Credits are held when you submit and reversed in full if the job fails or is cancelled.
How many Veo seconds does a plan actually cover? +
Hobby's 500 monthly credits buy 3 seconds, which is less than one 4-second clip at 640 credits. Pro's 2,000 buy 12 seconds, Studio's 6,000 about 37, and Agency's 10,000 per seat about 62. Top-up packs never expire and come in 600, 2,500 and 5,000 credit sizes.
Why did my square or 21:9 request fail? +
Veo 3.1 accepts 16:9 and 9:16 and nothing else. On text-to-video and animate, Playcut rejects the rest with a 400 before credits are held, so the failed attempt cost you nothing. For 21:9, use Seedance 2.0 or HappyHorse 1.1; 4:5 is HappyHorse only. Five of the six other video models take 1:1 — HappyHorse, Wan 2.7, both Grok Imagine models and Seedance 2.0 — while Gemini Omni Flash is 16:9 and 9:16 like Veo.
Why can I only pick 4, 6 or 8 seconds? +
Duration on Veo is a fixed enum rather than a range, and anything premium collapses it to 8. 1080p, 4K, reference mode, interpolation and extension all require exactly 8 seconds. Only 720p text-to-video and 720p animate leave the 4 and 6 second options open.
Can I get the same clip twice by reusing a seed? +
No. Playcut accepts a seed and records it on the asset for provenance, but it is never forwarded to Google, so reruns are not reproducible. Lock a look with reference images or an interpolation start frame instead.
Can I call Veo 3.1 from my own code or from Claude? +
Yes. All five modes are public tools reachable over REST at POST /api/v1/tools/{name}/execute with an Authorization: Bearer pk_live_ key, and the same five tool names are exposed over MCP to Claude, Cursor and ChatGPT. Submit the job, then poll get-task-status.
Ready to make something cinematic?
Pick a plan. Hobby starts with a 7-day trial — cancel within 7 days at no charge.