AI Video Generation Cost: Real Per-Second Rates [2026]

One second of AI video costs between $0.06 and $0.70 as of August 7, 2026, depending on which model you pick and what resolution you ask for. That is an eleven-fold spread for output that all gets described as "AI video" in the marketing.

The table below is the whole answer. We pulled these rates from official rate cards and reseller listings on August 7, 2026, three days after FLUX 3 Video went generally available and reset the price floor for long single-take clips.

What Does a Second of Output Cost Across the Major Models?

Prices are per second of finished output on pay-as-you-go API access. The last column shows a 20-second deliverable, which is the length most social ad formats want. Where a model cannot produce 20 seconds in one pass, the cost assumes stitching consecutive takes and the take count is labeled.

Two things fall out of that table immediately. First, the cheapest deliverable 20 seconds is a three-way tie on sticker price: Sora 2, Gemini Omni Flash, and Veo 3.1 Fast all land near $2.00 to $2.40, and all three make you stitch, since Sora 2 caps a single pass at 12 seconds, Omni Flash at 10, and Veo Fast at 8. The 25-second single pass belongs to Sora 2 Pro, at Pro prices. Second, FLUX 3 Video is the only model on the list that will hand you 20 continuous seconds at 1080p with synced audio for under six dollars in a single API call.

What Does a 20-Second Ad Actually Cost End to End?

A 20-second spot with three shots, a voiceover line, and one product beat is the most common brief we see. Here is the real math on four models, assuming you land the shot in a small number of attempts rather than one.

FLUX 3 Video, Full HD, single take. One 20-second generation is $5.80 with audio baked in. Say four attempts before you like one: $23.20. Now run the draft ladder instead. Six draft passes at $0.06 per second cost $7.20 total, you pick the composition that works, then spend $5.80 on one Full HD finish. Total $13.00, and you saw six variations instead of four.

Veo 3.1 Standard, three stitched takes. Twenty-four seconds of output (three 8-second passes) at $0.40 per second is $9.60 for one complete pass at the spot. Three attempts on each shot puts you at $28.80 before editing, and you still have to match color and audio across three separately generated clips.

Gemini Omni Flash, two stitched takes. Twenty seconds at $0.10 per second is $2.00. Cheapest credible route on this list. The catch is 720p at 24fps, which is fine for a vertical social cut and thin for anything that ends up on a landing page hero.

MiniMax H3 at 2K, two stitched takes. $2.60 at MiniMax's own rate, or $5.20 through fal.ai for the identical model. Same weights, same prompt, double the invoice. This is the single most avoidable cost mistake in the category right now.

Normalize the attempts before you compare those numbers. One complete pass at the 20-second spot is $2.00 on Gemini Omni Flash and $9.60 on Veo 3.1 Standard, a 4.8x spread driven by rate and resolution. Retakes then multiply both ends identically: at three attempts per shot the same two routes become $6.00 and $28.80, still 4.8x. So the honest summary is a roughly 5x spread across credible routes set by tier and resolution, with your retake rate scaling whichever route you picked rather than widening the gap.

What We Found Cross-Referencing Eight Rate Cards

We pulled published rates for every major video model on August 7, 2026 from eight sources: Black Forest Labs, Google (Gemini API and Vertex AI), OpenAI, MiniMax, Kuaishou, Runway, ByteDance, and fal.ai. Three findings that do not show up in any single provider's pricing page.

The same model at the same resolution is priced 2x apart. At 2K output, MiniMax H3 runs $0.13 per second on MiniMax's own rate card, roughly $0.1625 at Segmind, and roughly $0.26 at fal.ai. Same weights, same output tier, double the invoice at the top of the range. You are paying for whoever's GPU pool you land on. MiniMax also lists a $0.09 per second 768p tier in closed beta, but that is a lower output resolution, not a better price on the same thing, so it does not belong in the spread. If you are calling H3 through the first aggregator you found, check the others.

Nobody bills in the same unit. Black Forest Labs and Google bill per output second. ByteDance bills Seedance 2.5 per token on BytePlus ModelArk ($10.70 per million tokens without video input, $6.40 with), which means your cost moves with resolution and duration in a way you cannot predict from a rate card. Runway and Kling bill in credits, and Runway's API credits are bought separately from subscription credits. Comparing "per second" across the field requires normalizing three incompatible billing units, which is why most comparison posts quietly leave models out.

Max take length, not price, decides your bill. Six of the models above cannot produce 20 continuous seconds. Veo 3.1 caps at 8 seconds per pass, Gemini Omni Flash at 10, Runway Gen-4.5 at 10, Sora 2 at 12 (the 25-second pass is Sora 2 Pro only), MiniMax H3 at 15. Each stitch point is a place where lighting, character appearance, and audio ambience can drift, and every fix is another paid generation. FLUX 3 Video at 20 seconds and Seedance 2.5 at 30 seconds are the only two models that sidestep the problem, and Seedance 2.5 had no published per-second dollar rate we could buy at directly on August 7, 2026. The one fixed public rate we know of runs through a platform: Seedance 2.5 on Morphed is priced at 34 credits per second at 480p and 73 at 720p, audio included, for takes up to 30 seconds. That makes a full 30-second draft 1,020 credits and a 720p finish 2,190, with no token math involved.

The recent change most posts have not caught up with. OpenAI's deprecations documentation lists the Videos API along with sora-2 and sora-2-pro for removal on September 24, 2026. If you are building a cost model around Sora 2's $0.10 per second, check that date before you commit a pipeline to it.

Why the Sticker Rate Is Not What You Pay

Per-second pricing describes the cost of one successful generation. Nobody gets one successful generation.

Retakes are the dominant cost line in every real project, and the multiplier is entirely a function of how well you write the prompt. We are not going to hand you a made-up "average retakes per clip" number because we do not have clean first-party data on it. What we can tell you is the structure of the fix.

Use draft tiers as a search step, not a quality step. FLUX 3 Video's $0.06 per second draft mode exists so you can burn ten compositions for the price of two finished seconds. Draft output is not deliverable. It is a way to find out whether your camera move, your subject blocking, and your pacing work before you pay 4.8x more per second to render them properly. On Morphed the same pattern runs as 720p drafts at 25 credits per second, then a 1080p finish at 42 credits per second once the shot is locked.

Prompt quality is the biggest single cost lever. A prompt that specifies camera distance, lens behavior, lighting direction, and what the subject does in the last three seconds lands far more often than a prompt describing a mood. For model-specific structure, we wrote FLUX 3 Video prompts and Hailuo 3 prompts. Both cover the syntax each model actually rewards, which is not the same syntax.

Audio changes the arithmetic. FLUX 3 Video includes native audio in its listed rate. Google prices Veo 3.1 higher with audio than without, and Kling 3.0 publishes separate credit rates for silent and native-audio generations. If you are comparing a silent tier on one model against an audio-inclusive tier on another, you are not comparing the same product. Generating video silently and adding voiceover later means a separate tool, a separate bill, and a sync pass.

Should You Call the API Directly or Use a Platform?

API-direct is cheaper per second and more expensive per hour. That is the whole tradeoff, stated honestly.

Going direct to Black Forest Labs or the Gemini API gets you the rate on the card with no markup. What you take on: API key management across every provider you want to use, a polling loop for asynchronous jobs, retry and backoff handling when a generation fails, storage and CDN for outputs, a queue for batch runs, and a UI if anyone other than you needs to use it. For a team shipping one video a week, that build is not worth it. For a team generating a thousand clips a month on one model, it usually is.

Platforms sit on the other side. You pay a markup and you get a single credit balance across many models, a comparison view, retries handled for you, and generated files that live somewhere you can find them next month. Morphed runs flat credit rates across FLUX 3 Video, MiniMax H3, and the rest of the field in one workspace, so switching models is a dropdown rather than a new integration. It is a paid product with no free tier, which is the honest framing: you are trading dollars for not building infrastructure.

The middle path most teams land on: platform for exploration and client-facing work, direct API for the one high-volume workflow that has stabilized.

How Do Morphed Credits Map to Clip Length?

Morphed bills in credits per second of output, with the rate set by model and resolution. No per-generation surcharge for audio.

The 720p to 1080p step on FLUX 3 Video is a 1.68x credit increase, which is why the draft-then-finish pattern matters. Running five 720p explorations at 20 seconds costs 2,500 credits, and one 1080p finish costs 840. Running five 1080p explorations costs 4,200 credits for the same amount of learning.

MiniMax H3 caps at 15 seconds, so a 20-second deliverable means two generations and an edit. If your concept depends on reference images for character or product consistency, MiniMax H3 reference handles that path. For a head-to-head on which of the two leading models suits which shot, see Hailuo 3 vs Veo 3.

When Is Filming It Yourself Cheaper Than Generating It?

Sometimes the correct spend on AI video is zero. Three cases where we would tell you not to bother.

One hero shot you own the props for. A product on a table with window light, shot on a current phone, costs nothing and takes ten minutes including the retakes. Generating the same shot at 1080p with audio costs $5.80 on FLUX 3 Video and will not match your actual product's label, finish, or proportions without reference work. If the object is in the room with you, film it.

Minute-long spots with heavy continuity. A 60-second narrative with six shots, a recurring character, and matched lighting means six generations minimum, plus every retake, plus color matching across clips that were never rendered together. At Veo 3.1 Standard rates that is $24.00 for one clean pass at 60 seconds of output before a single retake, and the continuity work in the edit is where the real hours go. Chained AI takes get expensive faster than most people budget for.

Anything requiring a specific real person or real location. Reference-driven models have gotten good, and they still cannot deliver your actual storefront or your actual founder reliably enough for a brand spot. The cost is not the generations, it is the six rounds of client feedback about why the face is slightly wrong.

Where AI video does win on cost, clearly: shots that need a crane, a drone, a crowd, weather you do not have, or a location you cannot access. And variant testing, where you need the same concept twelve ways to find out which hook holds attention. Twelve 10-second variants on Gemini Omni Flash costs $12.00 and one afternoon. A shoot does not.

What to Do With This

Pick on max take length first, then price. If your deliverable is 20 seconds and the model caps at 8, the stitching work will cost more than the rate difference ever saves you.

Then run drafts before finals, always. The $0.06 per second draft tier on FLUX 3 Video and the 720p tier on Morphed exist for exactly this, and skipping them is the most common way teams turn a $6 clip into a $30 one.

For model selection beyond price, our best AI video generators roundup covers output quality and the best text-to-video AI generators comparison scores prompt adherence and motion. To generate across FLUX 3 Video, MiniMax H3, and the rest from one credit balance, start on Morphed.

All prices in this post were verified on August 7, 2026 and move often. Check the provider's rate card before committing a budget.