The cost math — when AI video is viable and when it bankrupts you (step 6/6) · ai video generation

Checkpoint

One last thing before we move on. Same surface as a write step — but the lesson doesn't complete until this passes.

Final drill. Synthesize price + retake + quality into a router: route_to_model(budget, quality_floor, retake_tolerance) that takes:

budget (float): max dollars per finished minute the user can spend.
quality_floor (str): minimum acceptable quality, one of "ai_flavored", "client_grade", "premium".
retake_tolerance (float): the retake multiplier the user expects to hit.

Each model has a quality_tier per this table:

Model	quality_tier	$/sec
Hailuo 02 (768p)	`ai_flavored`	0.045
Veo 3.1 Lite (720p)	`ai_flavored`	0.05
Kling 3.0 standard	`client_grade`	0.084
Sora 2 (720p)	`client_grade`	0.10
Sora 2 Pro (720p)	`premium`	0.30
Veo 3 (with audio)	`premium`	0.40
Sora 2 Pro (1024p)	`premium`	0.50

Quality ordering (each tier accepts higher tiers as substitutes): ai_flavored < client_grade < premium. So if quality_floor == "client_grade", both client_grade and premium models are eligible.

Algorithm:

Filter models down to those whose quality_tier is >= quality_floor.
Compute each eligible model's $/min real = price_per_sec × 60 × retake_tolerance.
Keep only models where $/min real <= budget.
From the remaining set, return the CHEAPEST one (lowest $/min real). On a tie, prefer the model that appears EARLIER in the PRICES dict order.
If no model fits, return "NO MATCH".

Five scenarios run. Expected output:

$10 + ai_flavored + 2.5x       -> Hailuo 02 (768p)
$20 + client_grade + 3.0x      -> Kling 3.0 standard
$50 + premium + 1.5x           -> Sora 2 Pro (720p)
$200 + premium + 4.0x          -> Sora 2 Pro (720p)
$2 + client_grade + 5.0x       -> NO MATCH