The cost math — when AI video is viable and when it bankrupts you — step 2 of 6
The retake rate — why every model is 2-10x more expensive than the table
The cost-per-second table from step 1 is the floor. The ceiling depends on retake rate — how many generations you actually have to do to land one usable clip.
Nobody talks about this in the marketing material. Everyone who has shipped real AI video work knows the number.
What retake rate is
If you generate one 5-second Sora 2 clip and use it as-is, your retake rate is 1.0. You spent $0.50, you have $0.50 of finished video.
If you generate ten 5-second clips and pick the best one, your retake rate is 10.0. You spent $5.00, you have $0.50 of finished video. The other $4.50 went into "I generated nine that didn't work."
This number is almost never 1.0 in real production. Even the best AI video operators average 2-5x in image-to-video mode and 5-10x in text-to-video mode for premium client work.
Why retake rate exists
Four reasons every generation might not be usable:
- The model misinterpreted the prompt. You asked for a medium close-up; you got a wide. You asked for an orbit; you got a pan.
- The physics broke. Fingers melt. Hair phases through a shoulder. A glass of water freezes mid-pour.
- The character drifted. In a sequence, the same character needs to look like the same character. Sometimes the model gives you almost-the-same-person.
- The motion was wrong-paced. You asked for a slow push-in; you got a fast one (or the move didn't happen at all).
The heavyweight models (Sora 2, Veo 3) fail less often. The cheap-tier models (Hailuo, Vidu) fail more often. Higgsfield lowers retake rate on camera-control failures specifically (its whole product), but doesn't help with content failures.
Working retake-rate estimates (mid-2026)
These are honest field numbers, not vendor marketing:
| Use | Workflow | Retake rate |
|---|---|---|
| Quick concept draft | T2V | 3.0x |
| Usable B-roll | T2V | 5.0x |
| Hero / brand shot | T2V | 10.0x |
| Quick concept draft | I2V | 1.5x |
| Usable B-roll | I2V | 2.5x |
| Hero / brand shot | I2V | 4.0x |
| Style transfer | V2V | 2.0x |
Three patterns worth internalizing:
- I2V cuts retake rate roughly in half vs T2V at every tier.
- Hero shots cost 4x more than B-roll in real spend, because the bar is higher.
- Style transfer (V2V) is cheap on retakes because the source video locks in the motion.
The real cost formula
For a finished minute of video:
real_cost_per_minute = price_per_sec × 60 × retake_rate
Run the math for "60 seconds of hero brand shots, image-to-video, on each model":
| Model | $/sec | Retake (I2V hero) | $/finished minute |
|---|---|---|---|
| Hailuo 02 | $0.045 | 4.0 | $10.80 |
| Veo 3.1 Lite (1080p) | $0.08 | 4.0 | $19.20 |
| Kling 3.0 standard | $0.084 | 4.0 | $20.16 |
| Sora 2 (720p) | $0.10 | 4.0 | $24.00 |
| Vidu Q3 | $0.07 | 4.0 | $16.80 |
| Sora 2 Pro (1024p) | $0.50 | 4.0 | $120.00 |
| Veo 3 with audio | $0.40 | 4.0 | $96.00 |
A 60-second hero piece on Veo 3 with audio at I2V retake rate costs $96 in raw API spend. The Sora 2 Pro path is $120. The Hailuo path is $10.80 — but the frame quality won't survive a Super Bowl spot.
The retake rate alone is the difference between "viable" and "not viable" for a lot of project ideas. Step 4 walks through a 60-second promo end-to-end across each model.
Three honest lessons
- Always plan budget against
price × duration × retake. Never against the headline price alone. - Retake rate is the single biggest variable you control. I2V vs T2V is a 2x reduction. Hand-crafted shot lists (lesson 02) reduce retakes by another 30-50% on top of that.
- Cheap models look cheaper than they are when the retake rate is poor. A model that fails 8x to deliver a usable shot is not cheaper than a model that fails 2x at higher per-second pricing. The retake math is what matters.
When somebody pitches you "AI video at $0.05/sec," ask: "What's the retake rate?" If they don't have a number, they haven't shipped.
The cost math — when AI video is viable and when it bankrupts you — step 2 of 6
The retake rate — why every model is 2-10x more expensive than the table
The cost-per-second table from step 1 is the floor. The ceiling depends on retake rate — how many generations you actually have to do to land one usable clip.
Nobody talks about this in the marketing material. Everyone who has shipped real AI video work knows the number.
What retake rate is
If you generate one 5-second Sora 2 clip and use it as-is, your retake rate is 1.0. You spent $0.50, you have $0.50 of finished video.
If you generate ten 5-second clips and pick the best one, your retake rate is 10.0. You spent $5.00, you have $0.50 of finished video. The other $4.50 went into "I generated nine that didn't work."
This number is almost never 1.0 in real production. Even the best AI video operators average 2-5x in image-to-video mode and 5-10x in text-to-video mode for premium client work.
Why retake rate exists
Four reasons every generation might not be usable:
- The model misinterpreted the prompt. You asked for a medium close-up; you got a wide. You asked for an orbit; you got a pan.
- The physics broke. Fingers melt. Hair phases through a shoulder. A glass of water freezes mid-pour.
- The character drifted. In a sequence, the same character needs to look like the same character. Sometimes the model gives you almost-the-same-person.
- The motion was wrong-paced. You asked for a slow push-in; you got a fast one (or the move didn't happen at all).
The heavyweight models (Sora 2, Veo 3) fail less often. The cheap-tier models (Hailuo, Vidu) fail more often. Higgsfield lowers retake rate on camera-control failures specifically (its whole product), but doesn't help with content failures.
Working retake-rate estimates (mid-2026)
These are honest field numbers, not vendor marketing:
| Use | Workflow | Retake rate |
|---|---|---|
| Quick concept draft | T2V | 3.0x |
| Usable B-roll | T2V | 5.0x |
| Hero / brand shot | T2V | 10.0x |
| Quick concept draft | I2V | 1.5x |
| Usable B-roll | I2V | 2.5x |
| Hero / brand shot | I2V | 4.0x |
| Style transfer | V2V | 2.0x |
Three patterns worth internalizing:
- I2V cuts retake rate roughly in half vs T2V at every tier.
- Hero shots cost 4x more than B-roll in real spend, because the bar is higher.
- Style transfer (V2V) is cheap on retakes because the source video locks in the motion.
The real cost formula
For a finished minute of video:
real_cost_per_minute = price_per_sec × 60 × retake_rate
Run the math for "60 seconds of hero brand shots, image-to-video, on each model":
| Model | $/sec | Retake (I2V hero) | $/finished minute |
|---|---|---|---|
| Hailuo 02 | $0.045 | 4.0 | $10.80 |
| Veo 3.1 Lite (1080p) | $0.08 | 4.0 | $19.20 |
| Kling 3.0 standard | $0.084 | 4.0 | $20.16 |
| Sora 2 (720p) | $0.10 | 4.0 | $24.00 |
| Vidu Q3 | $0.07 | 4.0 | $16.80 |
| Sora 2 Pro (1024p) | $0.50 | 4.0 | $120.00 |
| Veo 3 with audio | $0.40 | 4.0 | $96.00 |
A 60-second hero piece on Veo 3 with audio at I2V retake rate costs $96 in raw API spend. The Sora 2 Pro path is $120. The Hailuo path is $10.80 — but the frame quality won't survive a Super Bowl spot.
The retake rate alone is the difference between "viable" and "not viable" for a lot of project ideas. Step 4 walks through a 60-second promo end-to-end across each model.
Three honest lessons
- Always plan budget against
price × duration × retake. Never against the headline price alone. - Retake rate is the single biggest variable you control. I2V vs T2V is a 2x reduction. Hand-crafted shot lists (lesson 02) reduce retakes by another 30-50% on top of that.
- Cheap models look cheaper than they are when the retake rate is poor. A model that fails 8x to deliver a usable shot is not cheaper than a model that fails 2x at higher per-second pricing. The retake math is what matters.
When somebody pitches you "AI video at $0.05/sec," ask: "What's the retake rate?" If they don't have a number, they haven't shipped.