💸 What AI Video Really Costs in 2026: Per-Second to Per-Shot

Every pricing comparison you'll find lists a per-second number and calls it done: Kling here, Sora there, Veo at the top. That number is real, but it's not the number that determines your actual bill. We've shipped enough AI video to know the metric that matters is cost per finished shot — and that depends as much on how many takes a model needs as on its sticker price.

By the numbers

• Published per-second API pricing spans roughly $0.05/sec on budget-tier models up to $0.75/sec on premium native-audio models — trackers disagree on exact mid-tier numbers, so treat any single figure as a snapshot, not gospel
• Our own Seedance vs Veo comparison logs Seedance's volume-first pricing against Veo's native-audio premium — the two ends of the spectrum in daily use, not just on a price sheet
• In our production experience, character-driven or precise-motion shots typically need 3-5 takes per usable keeper — a number we track ourselves and reference across our cost-math pieces, not a vendor-published figure

The number nobody's pricing page shows you

Cost per finished shot = (price per second × clip length) × average takes needed. That last multiplier is the whole game. A model priced at $0.10/sec that needs five attempts to nail a specific camera move costs more, in practice, than a $0.15/sec model with tighter prompt adherence that gets there in two. Model pages advertise the first number and never mention the second — which is exactly why cheaper-on-paper models don't always win in a real production budget.

How we picked and tested this

We didn't run a fresh side-by-side render bake-off for this piece — we cross-referenced multiple public API pricing trackers (they disagree by 20-40% on mid-tier models, which is itself a useful signal about how fast this market moves) against our own production log: the models we actually pay for weekly, and roughly how many attempts each one needs per finished shot in our workflows. Where a number below is a hard published price, we say so; where it's our own operating average, we say that instead.

Where the money actually goes

1. The draft-to-finish multiplier. We render every shot at a lower resolution or shorter duration first to confirm the take before spending on a final-quality generation — the same economics behind running Wan 2.2 locally for drafts even when the finished shot ships from a cloud model. Skipping the draft step is the single fastest way to overspend.

2. Retakes on precise motion or dialogue. Anything requiring a specific gesture, camera timing, or lip-synced line costs more in practice than ambient b-roll, because it needs more attempts regardless of the model's advertised price.

3. Native audio as a line-item, not a footnote. Models with built-in audio generation (Veo 3.1) sit at the top of per-second pricing for a reason — you're paying for a feature that would otherwise mean a separate voiceover and sound-design pass. Whether that premium is worth it depends on whether you'd have paid for those passes anyway.

4. Subscription vs pay-per-generation. Credit-bundle platforms fold multiple models into one monthly allowance, converting variable per-second pricing into a flat, predictable cost — useful for steady output, a poor deal if your usage is bursty or seasonal.

The cut-list — where we trim first

• Cut resolution before you cut model quality. A 480p draft in the right model beats a 1080p final in the wrong one, and it's cheaper to iterate on.
• Cut speculative generations on shots you haven't scripted precisely. Vague prompts burn budget on takes you were never going to keep — write the shot down to subject, camera, lighting and setting first, the same discipline that saves credits in our prompt libraries.
• Cut native-audio models for shots where you're adding a separate voiceover anyway. You're paying for a feature you're about to override.
• Cut subscription tiers you're not using near capacity. Reconcile actual monthly usage against the plan every quarter — credit platforms are easy to over-provision on.

Pros and cons of optimizing for cheapest-per-second

Pros

• Real savings on high-volume, forgiving content — b-roll, ambient shots, anything without precise motion requirements
• Forces a draft-first discipline that improves workflow even beyond the cost savings
• Local models (Wan 2.2 and similar) become genuinely competitive once volume is high enough — see our local vs cloud cost math

Cons

• Chasing the cheapest sticker price on complex shots often costs more once retakes are counted
• Cheaper models frequently mean weaker prompt adherence, which is the actual driver of the retake problem
• Time spent iterating has its own cost that a per-second number never captures

What we didn't cover here

We're not publishing an exact price table claiming precision down to the cent — the trackers we cross-checked disagree with each other by enough that any single number would be false confidence dressed up as data. If you need current, exact pricing for a specific model, go to that model's own pricing page; we'd rather point you there than repeat a number that's stale by the time you read it.

Sources and further reading

Pricing figures referenced above come from public API cost trackers as of July 2026 — treat every number as a snapshot and re-check before budgeting a real project. For the two models we use most, see our first-hand Seedance vs Veo comparison, and for the local-hardware side of the equation, our local vs cloud AI cost math walks through the break-even in more detail.