Guides
AI Video Cost Per Second Compared 2026: Every Major Model
Mid-2026 cost-per-second comparison of Sora 2, VEO 3.1, Runway Gen-4, Kling 3.0, Wan 2.7, PixVerse V6 and more. Real math, hidden fees, optimization tactics.
Cost-per-second is the only AI video metric that matters once you stop experimenting and start shipping. Every other comparison — quality, resolution, motion fidelity — folds back into this number, because budget defines what you can actually produce on a monthly basis. The problem is that no two AI video vendors price the same way. Sora 2 sells subscription quotas. VEO 3.1 sells per-second metered Vertex AI compute. Runway sells credits. Wan 2.7 sells nothing — you rent the GPU yourself.
This guide normalizes all of them to the same unit: dollars per finished second of usable output. "Finished" matters, because regenerations and discards drive your real per-second cost up by 3-8x depending on the model. All figures are approximate mid-2026 estimates and are subject to change.
Section 1: Current pricing tiers across the major models
Here is the headline pricing landscape as of mid-2026. Cost-per-finished-second figures assume a typical hit rate (1 in 4 to 1 in 8 generations becomes a keeper).
| Model | Access Route | Headline Price | Approx Cost / Finished Second |
|---|---|---|---|
| Sora 2 | ChatGPT Plus | ~$20/mo | ~$0.05-$0.10 (quota-limited) |
| Sora 2 | ChatGPT Pro | ~$200/mo | ~$0.30-$0.70 |
| VEO 3.1 | Gemini Advanced | ~$20/mo | ~$0.10-$0.20 (quota-limited) |
| VEO 3.1 1080p | Vertex AI metered | ~$0.35-$0.50/sec billed | ~$2.00-$2.50 |
| VEO 3.1 4K | Vertex AI metered | ~$0.75-$1.20/sec billed | ~$4.00-$5.50 |
| Runway Gen-4 | Standard plan | ~$15/mo entry | ~$0.30-$0.50 |
| Runway Gen-4 | Unlimited plan | ~$76/mo | ~$0.10-$0.25 |
| Kling 3.0 | Pro subscription | ~$10-$30/mo + credits | ~$0.15-$0.35 |
| PixVerse V6 | Credit packs | ~$10-$50 packs | ~$0.10-$0.20 |
| Hailuo | Subscription + credits | ~$15/mo entry | ~$0.15-$0.30 |
| LTXV2 | Open source / self-host | GPU rental only | ~$0.02-$0.08 |
| Wan 2.7 | Apache 2.0 / self-host | GPU rental only | ~$0.02-$0.08 |
The price range across the market spans nearly two orders of magnitude. Self-hosting Wan 2.7 on a rented H100 is roughly 100x cheaper per second than metered VEO 3.1 4K — but the engineering overhead is real, and the quality ceiling is lower.
Section 2: How cost per second plays out on real projects
Project A — TikTok creator, 60 short clips per month
Need: 60 clips × 10s = 600 seconds of finished output. Quality bar: short-form social, 1080p, casual.
- Sora 2 Plus at $20/mo: well within quota, ~$0.03/sec effective. Winner for solo creators.
- Kling 3.0 Pro at $30/mo: comfortable with credits, ~$0.05/sec.
- Runway Standard at $15/mo: credit caps make this tight; expect overage charges.
Project B — DTC brand, 12 polished 30s ads per month
Need: 12 × 30s = 360 seconds finished, 1080p, brand-consistent.
- Runway Unlimited at $76/mo: works well, ~$0.21/sec effective.
- Sora 2 Pro at $200/mo: easy headroom, ~$0.55/sec effective.
- VEO 3.1 on Vertex AI: roughly $720-$900/month, ~$2.00-$2.50/sec effective. Highest quality, highest cost.
Project C — Production studio, 4K cinematic short
Need: one 3-minute 4K piece = 180 seconds finished.
- VEO 3.1 4K on Vertex AI: ~$900-$1,000 for the render budget, plus extensions.
- Sora 2 Pro × 2 seats: ~$400/mo, but 4K is selective on Sora 2.
- Wan 2.7 self-hosted on rented H100s: ~$15-$50 in GPU time, but engineering effort is substantial.
Project D — Agency, 200 client clips per month at mixed quality
A blended pipeline wins here. Most agencies in mid-2026 run a tiered stack: Wan 2.7 for b-roll and filler (cheap, infinite), Kling 3.0 or PixVerse V6 for mid-tier work, and Sora 2 Pro or VEO 3.1 for hero shots. Effective blended cost lands around $0.15-$0.30 per finished second across 6,000 monthly seconds.
Section 3: Hidden costs that wreck per-second math
Regeneration burn. This is the single largest hidden cost. A model with a 1-in-3 hit rate at $0.10/sec rate card is really $0.30/sec in your wallet. A 1-in-8 hit rate (common for tightly art-directed work) is $0.80/sec.
Quota ceilings. Subscription tiers look cheap until you hit the cap. The marginal cost of "the next second past the cap" is often infinite — you cannot top up, you can only upgrade the whole plan.
Watermarks. Free and entry-tier plans almost always watermark. Removing the watermark requires a tier upgrade, even if your usage would fit within the lower plan.
Audio pipeline. Most "video generation" pricing excludes branded voice. Add an AI voice cloning subscription (~$22-$99/month for ElevenLabs tiers) or AI lipsync costs to your real per-second math.
Storage and egress. Generated clips have to live somewhere. Vendor-hosted libraries have retention limits. Self-hosting means S3 or R2 bills.
Engineering overhead on open models. Wan 2.7 is free in license terms but the engineer-hours to operate a stable pipeline are real. Expect $500-$2,000/month of equivalent engineering attention even on a small deployment.
Vertex AI region pricing. GCP regional rates differ. US-Central is typically cheapest.
Extension fees. Pushing a clip past the base duration burns extra seconds on most models, sometimes nonlinearly.
Section 4: True cost per output across formats
Video is not the only output that matters. Here is how the math extends to adjacent media you will likely produce alongside.
| Output | Best-in-class option | Approx cost per unit |
|---|---|---|
| 1 generated image (1024×1024) | Flux 1.2 Ultra, Midjourney v7, Ideogram 3 | ~$0.04-$0.10 |
| 1 generated image (4K upscale) | Same + upscaler | ~$0.10-$0.25 |
| 1 minute of generated voice | ElevenLabs v3 | ~$0.08-$0.30 |
| 1 minute of generated music | Suno v5.5, Lyria | ~$0.10-$0.40 |
| 1 minute of lipsync output | AI lipsync pipelines | ~$0.30-$1.00 |
| 1 second of 1080p AI video | Mid-tier models | ~$0.10-$0.50 |
| 1 second of 4K AI video | Premium models | ~$1.50-$5.50 |
A 30-second branded video with custom voice, music, and a hero shot routinely runs $5-$25 in fully-loaded generation cost — not the headline subscription price. Plan accordingly.
Section 5: How to optimize cost per second
Tier your model usage. Use the cheapest acceptable model for each shot. A blended pipeline easily cuts costs 60% vs running everything through your premium tier.
Storyboard before you generate. Every regeneration is real money. A 10-minute sketch session saves 3-5 wasted renders per shot.
Render lower, upscale separately. 1080p + AI upscale is often 40-50% cheaper than direct 4K rendering with comparable web-delivery quality.
Generate b-roll on cheap models. Use an AI b-roll generator running on Wan 2.7, PixVerse V6, or LTXV2 for establishing shots and filler.
Batch overnight when possible. Off-peak rendering avoids queue contention on subscription tiers and can unlock priority on metered services.
Negotiate at scale. Enterprise contracts on Vertex AI, Runway, and ElevenLabs include volume discounts of 15-40% over rate-card pricing.
Stitch short clips instead of extending. Three 20-second clips edited together usually beats one extended 60-second clip on both cost and creative control.
Section 6: Common per-second cost mistakes
Comparing headline subscriptions instead of effective cost. A $20 plan with a 50-generation cap is not cheaper than a $76 unlimited plan if your project needs 200 generations.
Ignoring regen rate. The model with the lowest rate card but a 1-in-10 hit rate often costs more per finished second than a "more expensive" model with a 1-in-3 hit rate.
Defaulting to the most famous model. Sora 2 and VEO 3.1 are not always the right answer. Kling 3.0, Wan 2.7, and PixVerse V6 each win specific use cases.
Forgetting the audio and lipsync stack. Voice and music routinely add 20-40% to a finished video's effective per-second cost.
Skipping the test render. Running a 5-second sample before committing to a 60-second render saves money on every model.
Not benchmarking quarterly. Pricing in this market moves every 60-90 days. Re-run your math at least quarterly. See our budget breakdown for 2026 AI content costs and the best AI video models roundup for current snapshots.
FAQ
What is the cheapest AI video model per finished second in 2026?
Self-hosted Wan 2.7 or LTXV2 on rented GPU time is the cheapest at roughly $0.02-$0.08 per finished second, but requires significant engineering effort. Among hosted options, Sora 2 Plus inside its quota is the cheapest at ~$0.05-$0.10 per finished second.
Why is VEO 3.1 on Vertex AI so much more expensive than on Gemini Advanced?
Gemini Advanced is a flat $20/month subscription with a soft quota. Vertex AI meters every generated second with no cap, and prices reflect production-grade infrastructure. The same model, very different billing model.
How do I calculate my regen rate?
Track total generations submitted vs finished clips shipped over a month. A 1-in-4 rate means your effective cost-per-second is 4x your rate card. Most teams underestimate this and overspend.
Do these prices include audio and music?
No. Voice, music, and lipsync are separate pipelines on most models. Budget for an additional 20-40% on top of video generation costs for a fully-produced finished video.
How often does AI video pricing change?
Roughly every 60-90 days across the market. OpenAI, Google, and Runway have all adjusted pricing or quotas multiple times since January 2026. Always verify current rates before committing to a budget.
Cost-per-finished-second is the only metric that survives contact with a real shipping schedule. If you want one platform that gives transparent per-generation pricing across Sora 2, VEO 3.1, Kling 3.0, Wan 2.7, Runway Gen-4 and the full 2026 model lineup — without juggling six separate subscriptions — try the Versely AI video generator, pair it with text-to-image generation, and assemble final pieces with the AI movie maker.