AI Video Cost per Second in 2026: What a 30-Second Clip Really Costs

Introduction

A client asked me for a thirty-second product video last month and I quoted the job based on the per-second rate of the model I planned to use. The invoice at the end of the week was four times my estimate.

Nothing went wrong, exactly. The model did what it said on the pricing page. What I had forgotten is that per-second pricing describes generation, not output. A thirty-second video is rarely thirty seconds of generation. It is thirty seconds of generation multiplied by every attempt that did not survive review, plus every shot that had to be split because the model caps clip length, plus the audio decision I did not make deliberately.

So this article is about the number that actually matters: cost per finished second of video you can publish. I have pulled the published per-second rates for nine models, applied the same realistic assumptions to each, and worked out what a thirty-second deliverable costs on each route. Some of the gaps are bigger than the quality gaps.

TL;DR

• Published rates in this market span roughly 13x, from about $0.05 to $0.79 per second at 1080p with audio. Model quality does not span 13x, so price and quality are not the same decision.
• Cheapest per finished thirty seconds: Veo 3.1 Lite at about $4.50, then LTX 2.3 Fast at $5.40. Both are legitimate routes, not compromises.
• Most expensive: Seedance 2.5 at about $71 for the same thirty-second deliverable. It earns that only when prompt adherence and reference consistency are the whole job.
• The three multipliers that wreck estimates are keeper rate (often 1 in 3), clip-length caps (an 8-second cap quadruples the generation count for a 30-second video), and the audio toggle (up to double the rate on some models).
• Wan 2.7 sits at about $13.50 per finished thirty seconds at a 1-in-3 keeper rate — the middle of the market, with the widest control set at that price.

Generate a Test Clip for Free →

The published rates

These are per-second API rates at 1080p with audio on, read in late September 2026. They come from provider documentation and a published cross-model cost comparison; my own worked examples below are derived from them, not measured on my own invoice.

Model	Rate (1080p, audio on)	Max clip	Notes
LTX 2.3 Fast	~$0.060/s	20s	Open weights available
Veo 3.1 Lite	~$0.050/s	8s	Cheapest route to clean output
Hailuo 2.3 Pro	~$0.071/s	6s at 1080p	Short clips only
Veo 3.1 Fast	~$0.143/s	8s	Balanced tier
Kling 2.6 Pro / O3 Pro	~$0.140/s	10–15s	Older Kling tiers
Wan 2.7	~$0.150/s	15s	Audio always on, no toggle
Kling V3 Pro	~$0.168/s	15s	4K options available
Wan 3.0 Standard	~$0.200/s	30s	Single-pass 30s
Veo 3.1 Standard	~$0.393/s	8s	Highest-fidelity Veo tier
Seedance 2.0 Standard	~$0.527/s	15s	Reference-heavy workflows
Seedance 2.5	~$0.791/s	30s	Most capable, most expensive

Two things jump out immediately. First, the spread is enormous — thirteen times between the cheapest and most expensive route. Second, the cheapest models are not the worst models. Veo 3.1 Lite and LTX 2.3 Fast both produce publishable footage; they just do it with fewer knobs and shorter clips.

Why your invoice is never the table price

Per-second pricing is honest about generation and silent about everything else. Four multipliers sit between the table and your bill.

1. Keeper rate

The number of generations it takes to get one clip you would actually publish. For a straightforward product shot on a stable model, it can be close to 1 in 2. For anything with a face, a specific motion, or text on screen, 1 in 3 to 1 in 5 is normal. That multiplier applies to every second you generate.

2. Clip-length caps

If the model caps at 8 seconds and your deliverable is 30 seconds, you are generating at least four clips. Four clips means four separate keeper decisions, and each one has its own failure rate. This is where Veo's 8-second cap stops being a footnote and starts being a line item.

If you want output today, start here: Launch Wan 3.0 Now →

3. The audio switch

Several models charge roughly double when audio is on. Seedance 1.5 Pro is about $0.113/s with audio against $0.056/s silent. Veo 3.1 Standard is about $0.393/s with audio against $0.196/s. Kling V3 Pro is about $0.168/s against $0.112/s. If you plan to lay your own soundtrack, leaving audio on is a silent surcharge.

Wan 2.7 is the exception in the other direction: it returns audio and does not offer a toggle. If you wanted silence, you are paying for audio you will discard — worth knowing before you build a cost model around it.

4. Resolution

Every rate above is 1080p. Dropping to 720p is often 30–50% cheaper; going to 4K can double it. For social placements where the file gets recompressed anyway, 720p drafts are the rational choice.

What a finished thirty seconds actually costs

Here is the calculation nobody publishes. Assumptions, stated plainly: a thirty-second deliverable, a 1-in-3 keeper rate (three generations per usable second, which is realistic for anything with motion), and the clip split shown for each model based on its cap.

Model	Split for 30s	Generation needed	Cost per finished 30s
Veo 3.1 Lite	4 × 7.5s	90s	$4.50
LTX 2.3 Fast	3 × 10s	90s	$5.40
Hailuo 2.3 Pro	5 × 6s	90s	$6.43
Wan 2.7	2 × 15s	90s	$13.50
Kling V3 Pro	2 × 15s	90s	$15.12
Wan 3.0 Standard	1 × 30s	90s	$18.00
Veo 3.1 Standard	4 × 7.5s	90s	$35.36
Seedance 2.0 Standard	2 × 15s	90s	$47.46
Seedance 2.5	1 × 30s	90s	$71.18

Read the column on the right, not the rate column. A model at $0.20/s that gives you 30 seconds in one pass costs less per finished video than a model at $0.15/s that caps at 10 seconds, because the second one forces three keeper decisions instead of one.

Two consequences worth internalising:

• The clip-length cap is often worth more than the per-second rate. A 30-second single-pass model at twice the rate can still win, because it removes generation rounds and reduces the chance of a continuity mismatch between clips.
• The keeper rate dominates everything else. Move from 1-in-3 to 1-in-5 and every number in that table grows by two-thirds. This is why experienced prompting is the highest-return cost optimisation available — better than switching models, better than hunting for the cheapest endpoint.

Test Wan 2.7 at 1080p →

The cost of chasing quality

The expensive models are not overpriced for what they do. They are overpriced for most jobs.

Seedance 2.5 at roughly $0.79 per second costs about sixteen times Veo 3.1 Lite per finished thirty seconds. What you get for that: the strongest prompt adherence in the field, up to thirty image references, and single-pass thirty-second clips. If your job is "this exact product must look identical across eight shots with brand-accurate labels", that is not a luxury — a cheaper model that drifts costs more in retouching than the generation premium.

If your job is "a generic lifestyle shot that feels like our brand", the premium buys you nothing you can see at 1080p on a phone screen.

The decision rule I use:

Ready to try it yourself? Try Wan 3.0 Free →

Job type	Model tier that fits	Why
Social cuts, high volume	Veo 3.1 Lite, LTX 2.3 Fast	$5 per finished clip; imperfections vanish after compression
Product demos with brand assets	Wan 2.7, Wan 3.0	Reference control and 30s single pass at mid-market rates
Hero spot, client-facing	Veo 3.1 Standard, Seedance 2.5	Pay for the shot that gets scrutinised
Multi-shot narrative with characters	Seedance 2.5, Kling V3 Pro	Consistency across shots is the expensive part
Internal drafts and pitch boards	720p on the cheapest tier you have	Nobody evaluates a mood board at 4K

The route that halves the bill: draft cheap, finish expensive

The most effective cost structure is not a cheaper model. It is a two-tier workflow.

1. Draft at the cheapest tier, at 720p, short. Five seconds is enough to judge whether a prompt is producing the composition and motion you want. Nothing about that judgement improves at 1080p.
2. Lock the prompt. Only when the draft looks right do you spend on the final render. Most of your failures happen before the prompt is locked, so most of your spend should happen on the cheap tier.
3. Finish once, at full resolution, on the model that fits the job. One expensive generation with a proven prompt is cheaper than five expensive explorations.

Run the numbers: five 5-second 720p drafts on a cheap tier cost less than a single 1080p 10-second premium generation. If drafting cuts your keeper rate at the top tier from 1-in-3 to 1-in-1.5, you have halved the expensive half of the bill for pennies.

The same logic applies to free tiers: they are the cheapest drafting environment that exists. If you want to spend nothing while you iterate, free Wan video generation is a reasonable place to burn attempts before you pay for a render. For a closer look at how it stacks up against other models, see Best Image to Video AI Tools in 2026. If you want to test it without installing anything, the free image-to-video generator works in the browser.

When local actually saves money

Open-weight models — Wan 2.2, LTX 2.3 — move the cost from per-second billing to hardware and setup. The marginal cost of a generation becomes electricity.

The break-even is not where people assume. You do not need many videos to justify it; you need repeatability. If you are generating a hundred test clips a week while tuning a prompt, local is dramatically cheaper. If you are producing two finished clips a month on a laptop, the setup hours never pay back and the API is cheaper in every sense.

The one thing local always buys, regardless of volume: no queue, no content policy surprises mid-project, and no per-generation bill to justify internally.

Five traps that inflate video budgets

1. Generating at 1080p to evaluate a prompt. You are paying for resolution you cannot use to make a decision.

2. Ignoring the clip cap until the end. Discovering the 8-second cap after you have designed a 30-second sequence means re-planning the whole piece.

3. Leaving audio on when you will mute it. Up to double the rate, silently.

Want to see the difference on your own footage? Start creating with Wan 3.0 →

4. Re-rolling instead of editing. A model with instruction-based editing lets you fix one element. Re-rolling re-randomises everything and resets your keeper rate.

5. Quoting clients at the table rate. The number on the pricing page is a generation cost. Your quote has to cover generations per keeper, plus the shots that get cut in review.

FAQ

How much does a 30-second AI video cost in 2026?

Between about $4.50 and $71 depending on the model, at a realistic 1-in-3 keeper rate and 1080p with audio. Veo 3.1 Lite and LTX 2.3 Fast sit at the bottom, Wan 2.7 in the middle at roughly $13.50, and Seedance 2.5 at the top.

Why is the per-second price not my actual cost?

Because per-second pricing charges for generation, not output. You generate multiple attempts per usable clip, and models with short clip caps require several generations to cover one deliverable. Both multipliers sit on top of the table rate.

Which AI video model is cheapest per second?

Skip the setup and test it in the browser: Experience Wan 3.0 Free →

Veo 3.1 Lite at about $0.05 per second with audio, followed by LTX 2.3 Fast at about $0.06. Both produce publishable output; they trade off maximum clip length and control rather than basic quality.

Is it cheaper to generate at 720p or 1080p?

720p is typically 30–50% cheaper and is the right choice for drafts and for social placements that will be recompressed. Render the final at 1080p only after the prompt is locked.

Does audio cost extra?

On several models, yes — roughly double the rate when audio is enabled. Seedance and Veo both price audio-on higher than silent. Wan 2.7 returns audio with no way to disable it, so it always includes that cost.

Are open-source models actually free to use?

The weights are free; the compute is not. Wan 2.2 and LTX 2.3 can run locally at zero marginal cost per generation, but they require a capable GPU and setup time. It pays off with high volume or a need for repeatability.

How do I cut AI video costs without losing quality?

Draft cheap at 720p with short clips, lock the prompt, then render once at full quality on the model that fits the deliverable. Editing an existing output is cheaper than re-rolling the prompt, and prompt skill is the single biggest lever on keeper rate.

The Bottom Line

The per-second price on a model's pricing page is the least useful number in this decision. What matters is the cost of a finished second — after keeper rate, clip caps, audio, and resolution have all been applied. On that basis the market ranges from about $4.50 to $71 for the same thirty seconds of deliverable, and most jobs belong in the bottom third of that range.

Start cheap, draft at low resolution, lock your prompt, and spend on one proper render. That workflow routinely costs a third of what a single-model, full-resolution approach burns — and it produces better footage, because it forces you to judge the shot before you pay for it.

Start Drafting for Free →

Related guides

• Wan 3.0 complete guide — the 30-second single-pass model behind the $18 row
• Wan 2.6 vs Wan 2.7 comparison — where the mid-market tiers actually differ
• Wan 2.5 vs Kling 3 — quality difference for the price difference
• Wan 3.0 vs Wan 2.7 — single-pass 30s versus cheaper per second

References

• fal.ai pricing — published per-second rates across video models
• Hedra: Seedance, Kling, Veo and MiniMax API pricing — 1080p comparison of ten-second clip costs
• fal.ai: Seedance 2.0 vs Veo 3.1 — tiered pricing including silent versus audio rates
• fal.ai Kling 3.0 — Kling tier pricing range
• Artificial Analysis text-to-video leaderboard — quality benchmarks referenced when comparing price to output