AI Video Generation Cost in 2026: Price per Second and per Usable Shot

Calculate direct generation cost from official per-second or per-clip pricing, while separating observed prices from unverified retry averages. AI video generation cost has two different layers: the listed charge for each generated second or clip, and the project cost of obtaining a shot that passes review. Use the provider’s current official rate to calculate candidate cost, then divide total generation spend by accepted shots. Never insert an “average” retry count unless it comes from your own production log. For the ai video generation cost decision in the “direct answer” stage, this is review note 1: retain the named evidence and do not generalize beyond this shot brief.

There is no defensible public industry average for retry count, acceptance rate, or savings caused by a 3D reference. Measure those values on your own shots before making a performance claim.

Start with the billing unit, not a headline price

OpenAI lists Sora 2 Pro API video by output second and separates 720p, 1024p, and 1080p rates. Google likewise separates Veo variants, audio modes, and resolutions. Runway publishes credits per second plus a dollar value per credit. Luma publishes fixed prices for specified clip lengths and resolutions. Those units are not interchangeable. A sound comparison first records model, mode, resolution, duration, audio, minimum charge, currency, and access channel. It then calculates only the configuration actually under consideration. A subscription price, API rate, and third-party hosted rate should never be merged into one universal cost-per-second figure. Save the source URL and access date beside every rate because pricing is mutable. For the ai video generation cost decision in the “Start with the billing unit, not a headline price” stage, this is review note 2: retain the named evidence and do not generalize beyond this shot brief.

Official 2026 price examples and their limits

At the July 21, 2026 evidence snapshot, OpenAI listed Sora 2 Pro at $0.30 per second for 720p, $0.50 for its 1024-class output, and $0.70 for 1080p. Google listed Veo 3.1 at $0.40 per second with audio and $0.20 without audio for the specified modes. Runway listed Gen-4.5 at 12 credits per second and credits at $0.01, which yields $0.12 per output second for that API configuration. Luma listed Ray3.2 five-second T2V/I2V clips at $0.15 for 540p, $0.30 for 720p, and $1.20 for 1080p. These are scoped observations, not a permanent market average. For the ai video generation cost decision in the “Official 2026 price examples and their limits” stage, this is review note 3: retain the named evidence and do not generalize beyond this shot brief.

Calculate candidate cost with scenario arithmetic

Scenario arithmetic is straightforward when the provider bills by output second: candidate cost equals duration multiplied by the official per-second rate. A five-second 720p Sora 2 Pro candidate is therefore 5 × $0.30 = $1.50 under the cited API rate. Ten such candidates would be 10 × $1.50 = $15.00. This example is a planning scenario, not a claim that creators normally need ten candidates. For fixed-clip pricing, use the listed clip price rather than forcing a linear rate. Also preserve minimum-generation charges; a nominal per-second rate can understate a very short request when the API applies a minimum. For the ai video generation cost decision in the “Calculate candidate cost with scenario arithmetic” stage, this is review note 4: retain the named evidence and do not generalize beyond this shot brief.

Move from generated seconds to usable-shot cost

A budget is not finished when it reports only the selected clip. Every generated candidate consumes billed output even when it fails composition, timing, identity, motion, policy, or edit-readiness checks. Track generated seconds, direct generation spend, accepted shot count, and rejection reason. Cost per usable shot equals total direct generation spend divided by accepted shots. If a sequence costs $48 in generation and produces four accepted shots, the measured result is $12 per usable shot for that sequence. It says nothing about another team, model, or genre. Keep labor, editing, upscaling, storage, and API failures in separate columns so inference cost remains auditable. For the ai video generation cost decision in the “Move from generated seconds to usable-shot cost” stage, this is review note 5: retain the named evidence and do not generalize beyond this shot brief.

Example workflow: budget a six-shot product sequence

Suppose a team plans six five-second product shots. It first selects a model configuration and copies the current official unit price into a dated cost sheet. It then creates a shot control row for each angle, with duration, format, camera behavior, subject action, and rejection rules. Before final generation, the team blocks the sequence with a storyboard or graybox. During generation, every candidate receives an ID and actual billed amount. At approval, the sheet totals all candidates rather than multiplying six final shots by one candidate price. This workflow exposes where spend accumulated: a difficult orbit shot, a continuity problem, or repeated style changes. For the ai video generation cost decision in the “Example workflow: budget a six-shot product sequence” stage, this is review note 6: retain the named evidence and do not generalize beyond this shot brief.

Use graybox review as a decision gate, not a savings promise

SEELE AI can help externalize framing, layout, subject trajectory, and timing in a graybox or previs pass before final video generation. The defensible claim is procedural: reviewers can inspect these decisions earlier. The unsupported claim would be that this necessarily reduces generation cost by a fixed percentage. To test value, run matched shots with the same model, duration, resolution, prompt assets, and acceptance criteria. Compare total generated seconds, accepted-shot cost, camera deviation, and review time. Until that experiment exists, describe the graybox as a control method rather than guaranteed savings. For the ai video generation cost decision in the “Use graybox review as a decision gate, not a savings promise” stage, this is review note 7: retain the named evidence and do not generalize beyond this shot brief.

Build a cost ledger that survives model changes

A useful ledger stores date, provider, model version, channel, billing unit, resolution, audio mode, duration requested, duration billed, candidate ID, shot ID, status, rejection code, and direct charge. Add human time only in a separate labor table. When prices change, preserve the historical row rather than rewriting it with today’s rate. This allows finance and creative teams to reproduce what a campaign actually cost. It also prevents a low headline price from hiding a model mismatch: the cheapest candidate is not economical if it cannot satisfy the camera, duration, or continuity requirement. For the ai video generation cost decision in the “Build a cost ledger that survives model changes” stage, this is review note 8: retain the named evidence and do not generalize beyond this shot brief.

Decision rules for choosing a model tier

Choose a model tier by shot requirement, not by assuming the premium option always wins. Use lower-cost exploration for broad visual ideation when exact continuity is unimportant. Reserve higher resolution, audio generation, or expensive models for shots whose acceptance criteria require those capabilities. Run a small calibrated sample before scaling a sequence. The decision should compare measured cost per accepted shot and failure categories, not just unit price. If one model costs twice as much per second but reaches the required camera behavior with fewer measured candidates on your test, that project-specific result may justify it; it still is not an industry rule. For the ai video generation cost decision in the “Decision rules for choosing a model tier” stage, this is review note 9: retain the named evidence and do not generalize beyond this shot brief.

Practical next steps in SEELE AI

Start with Greybox previs when spatial or camera decisions need review, move to the AI video generator when the shot package is approved, and use Storyboard-to-video when sequence and beat planning are the main uncertainty. Keep one shot ID across planning, generation, and acceptance so evidence remains connected. These tools support a controlled workflow; they do not guarantee model obedience, acceptance rate, or cost savings. For the ai video generation cost decision in the “Practical next steps in SEELE AI” stage, this is review note 10: retain the named evidence and do not generalize beyond this shot brief.

Related guides in this controlled-video series

• What Does One Usable AI Video Shot Really Cost?
• Text-to-Video vs 3D Reference: Which Gives More Shot Control?
• Controlled AI Video Generation: From Prompting to Explicit Shot Constraints
• Graybox Animation for AI Video: Block the Shot Before You Render

Operational measurement workflow

Use this ordered workflow to turn ai video generation cost into a reproducible production decision rather than a vague aspiration:

1. Name the shot's viewer-facing job, duration, format, and accountable approver before selecting a model.
2. Write separate constraints for framing, subject behavior, camera motion, event timing, continuity, and the final frame.
3. Choose the cheapest honest planning artifact that exposes those decisions, such as a control sheet, storyboard, graybox, or 3D camera path.
4. Approve the planning artifact before final generation, while clearly marking style, lighting, and performance choices that remain flexible.
5. Record every generated candidate with model, settings, billed unit, generated duration, direct charge where available, and a stable review identifier.
6. Review candidates against the written controls before judging general visual appeal; classify each rejection as structural, temporal, compositional, factual, policy-related, or aesthetic.
7. Accept the shot, revise only the responsible input, or escalate an unresolved creative choice. Preserve the receipt so later reports use observed data instead of remembered estimates.

For example, a five-second camera move should not be accepted merely because it looks cinematic. The reviewer checks the agreed start frame, endpoint, subject path, timing, and required edit handles. A second workflow might test a product reveal whose logo side and final-frame hold are mandatory. A third might compare a text-only brief with a 3D reference under fixed model settings. These are project tests, not proof of a universal retry count or savings rate. For the ai video generation cost decision in the “Operational measurement workflow” stage, this is review note 11: retain the named evidence and do not generalize beyond this shot brief.

The resulting record is useful beyond one generation. Producers can see which control failed, finance can separate direct inference from labor, and directors can decide whether a new candidate, a changed reference, or an edit is the appropriate next action. That is the practical value of an explicit workflow: it makes the next decision legible without pretending stochastic generation has become deterministic. For the ai video generation cost decision in the “Operational measurement workflow” stage, this is review note 12: retain the named evidence and do not generalize beyond this shot brief.

FAQ

What is the simplest formula for AI video generation cost?

For per-second billing, multiply billed output seconds by the official rate for the exact model, resolution, audio mode, and access channel. For fixed-clip billing, add the listed clip charges. Keep minimum charges, taxes, labor, editing, storage, and rejected candidates visible instead of hiding them in one blended number.

How do I calculate cost per usable AI video shot?

Add the direct generation charges for every candidate associated with the shot or sequence, including rejected outputs, and divide by the number of shots that passed the written acceptance criteria. Label the model, date, settings, and sample size so the result remains a project measurement rather than a supposed market benchmark. For the ai video generation cost decision in the “FAQ” stage, this is review note 13: retain the named evidence and do not generalize beyond this shot brief.

Is there an average number of retries for AI video?

No reliable public industry-wide retry average is established by the cited pricing pages or benchmark papers. Retry behavior depends on model, prompt, duration, shot complexity, references, policy filters, and acceptance standards. Record candidates and rejection reasons in your own production log, then report the measured distribution for that defined sample. For the ai video generation cost decision in the “FAQ” stage, this is review note 14: retain the named evidence and do not generalize beyond this shot brief.

Does 3D previs automatically reduce AI video cost?

Not automatically, and no fixed savings percentage should be claimed without a controlled comparison. Previs can make camera, layout, scale, and timing decisions inspectable before generation. Whether that changes accepted-shot cost must be measured with matched shots, fixed model settings, equivalent review criteria, and all candidate charges included.

How often should official prices be checked?

Check the official page when approving a budget and record the access date, because model names, modes, minimum charges, resolutions, and rates can change. For a long campaign, revalidate before each material purchase batch. Do not silently apply a current price to historical generations that were billed under a different schedule. For the ai video generation cost decision in the “FAQ” stage, this is review note 15: retain the named evidence and do not generalize beyond this shot brief.

Sources and claim boundaries

• OpenAI Sora 2 Pro API pricing
• Google Vertex AI generative media pricing
• Runway API pricing
• Luma API pricing

Sources were accessed for the July 21, 2026 evidence snapshot. Pricing statements are scoped to the cited official API configuration and may change. Research findings are scoped to the paper’s tested models, prompts, and metrics. Scenario arithmetic is labeled and must not be treated as an industry average.