CREDITS & PRICING • Framework updated · Jul 2026
AI Video Cost per Minute: Raw Duration vs. Usable-Clip Cost
Cost per minute feels like the natural way to compare AI video tools, yet almost none of them price that way. Most sell credits that become seconds only after you choose a model, a resolution, and a clip length, so the minute you can actually deliver often costs far more than the meter implies. This guide separates raw-duration cost from usable-clip cost and walks through a worked estimate you can adapt to any provider.
Why per-minute rarely appears on a price page
Runway and Luma both meter video in credits, not minutes. Runway's ladder runs Free ($0, 125 one-time credits), Standard ($12 a month, 625 credits), Pro ($28, 2,250 credits), and Max ($76, 9,500 credits), with Enterprise priced on request. Luma sells Plus, Pro, and Ultra plans carrying 10,000, 40,000, and 150,000 credits. Neither publishes a dollars-per-minute figure, because a minute only exists once you convert credits into seconds at a chosen model and resolution. OpenAI's Sora API was the rare product that billed straight per second, and it is now being retired. Treat any per-minute number as something you derive from a plan, not something printed on the page.
Raw-duration cost versus usable-clip cost
Two different numbers hide inside cost per minute. Raw-duration cost is the sticker math: the price of generating sixty seconds of output once, straight through. Usable-clip cost is what you actually pay to ship sixty deliverable seconds, after failed generations, resolution or watermark upgrades, and the clips you discard. The gap is usually wide. A raw second assumes every generation is a keeper at the resolution your project needs, which rarely holds. Models miss prompts, warp hands, or drift off-brand, so teams generate several candidates per shot and keep one. They also find that an entry-tier clip carries a watermark or caps below their delivery resolution, forcing a paid upgrade before the footage is truly usable. Budget against usable-clip cost, and treat raw-duration cost as the floor you can never beat.
A worked sixty-second scenario
Say you need a 60-second finished sequence. Start with the raw-duration floor: target seconds times the per-second rate. On Sora's API list pricing, sora-2-pro at 1080p is $0.70 per second, so 60 seconds of raw output lands near $42 before anything goes wrong. Now layer reality. Sora's documented generation lengths top out at 20 seconds, so a 60-second piece is at least three separate billable calls, not one. Then apply your keep rate: if you accept one clip for every A you generate, multiply the raw cost by A. At two attempts per keeper, that $42 floor becomes roughly $84 of generation spend. On a credit tool the same logic runs in credits: Runway lists Gen-4.5 at 60 credits per five seconds of video, which is 12 credits a second, so one 60-second raw pass is 720 credits, already more than Standard's entire 625-credit month.
What quietly inflates the real number
Four forces push usable-clip cost above the raw floor. Retries come first: every regeneration re-spends the full price or full credits, so a low keep rate multiplies the bill directly. Second is deliverability gating: Runway lists No watermarks and 4K Upscaling as features of its paid Standard plan and above, and Sora requires the pricier sora-2-pro model for 1080p exports, so a clip that meets spec often costs more than the cheapest clip the tool can make. Third is per-call duration caps: because each Sora generation tops out at 20 seconds, longer pieces are stitched from multiple billable runs. Fourth is perishable credits: Runway's Standard and Pro credits reset within 24 hours of the billing date and do not roll over, so any allotment you fail to spend before reset raises your effective cost per delivered minute.
Reading each provider's meter
Because the unit differs by provider, convert everything to your own cost per delivered second before comparing. On Runway, translate credits to seconds through the model you will actually use: Gen-4.5's 60 credits per five seconds works out to 12 credits a second, and other models carry other rates, so a plan that looks generous can be thin for your workflow. On Luma, price scales with both resolution and length: the pricing page lists Ray3.2 at 20 credits for a 5-second draft-resolution clip, 60 credits for 10 seconds at the same quality, and 400 credits for 5 seconds at 1080p, so neither length nor resolution scales your cost linearly. On Sora's API the rate is explicit but the product is sunsetting: OpenAI notified developers of the Videos API and Sora 2 deprecation on March 24, 2026, with removal on September 24, 2026, so treat its clean per-second pricing as a benchmark, not a platform to standardize on.
Building your own per-minute estimate
Turn the framework into a repeatable calculation. Pick the exact model and resolution your deliverable requires, then find that combination's current per-second or per-credit rate on the provider's page. Multiply by your target seconds for the raw-duration floor. Multiply again by your realistic attempts-per-keeper to reach expected generation spend, then add any tier upgrade needed to remove watermarks or hit your resolution. If the tool uses non-rolling monthly credits, divide your true output by the credits you can actually consume before reset, not the credits the plan grants. The result is a defensible cost per usable minute for your specific project, which is the only per-minute number worth quoting. Because rates, model names, and credit allotments change often, re-check each figure on the official page the day you commit.
Editorial note: This framework is general information, not a vendor endorsement. Check the current pricing, terms, and data-handling details directly with the provider before buying.