What would it cost to make Breaking Bad with AI?
A line-by-line estimate for one 47-minute episode using this year's public video generation prices. Generating the footage is only part of the budget; performance and continuity remain much harder to get right.
Last April I worked out what a minute of AI video cost across a few models. People keep asking a bigger version of that question: forget one-minute microdramas, what would a real hour-long drama cost to make with AI? I’ll use Breaking Bad as the reference, because everyone knows it and because it’s about as far from a microdrama as television gets. These are list prices and my own assumptions, not anything from my job.
The reference point. A Breaking Bad episode is about 47 minutes. Its production budget was widely reported at around three million dollars per episode. That buys writers, a cast of excellent actors, a crew, locations in and around Albuquerque, weeks of shooting, and months of editing, sound and color.
Generation cost. The per-second prices have moved a lot since last year. In China, where prices are lowest, a comparison published in April found costs from about 2.5 cents per second at the cheap end to about 19 cents per second for Seedance 2.0 on Jimeng’s top membership, after Jimeng raised prices three times over the past year. Kling 3.0 Omni came in around 6 cents per second. Western API pricing is generally higher.
47 minutes is 2,820 seconds of finished footage. A drama, with close-ups where performance matters, needs a lot of retakes. I’ll assume ten generations for every second kept, which is pessimistic for establishing shots and optimistic for emotional close-ups. That’s about 28,000 seconds of generation. At 19 cents per second, that’s a bit over $5,000. At Kling’s price, it’s closer to $1,700. Even if Western prices are five times higher, the generation bill is in the tens of thousands of dollars, not millions.
What else you need. Writing, which is people, and for something like Breaking Bad it’s most of the value. Character and production design, meaning reference sheets for every character, costume and location. Voice, which you could generate, but dialogue performance is exactly where generation is weakest. Editing, sound design, music and color. And the labor of directing the generations: choosing takes, fixing continuity, re-prompting. For a small skilled team working a few weeks per episode, call it tens of thousands to low hundreds of thousands of dollars.
So a rough total is somewhere between $50,000 and $300,000 per episode, depending on the team and the tools. That’s roughly two to ten percent of the reported budget.
What would that money actually buy? Something with the look of prestige television, and very little of what made Breaking Bad what it is. What makes that show great is performance. Bryan Cranston’s face in the moment Walter decides something, Aaron Paul’s timing, silences held exactly as long as they should be. Current models can produce a convincing face making an expression. They can’t reliably produce a performance that develops across a scene, reacts to the other actor, and stays consistent with who the character has become over forty episodes. Continuity at that scale, the same face, voice and mannerisms across hours, is also still unsolved, even with Seedance 2.5’s 30-second shots, which arrived this summer.
By this estimate, generating enough footage for a Breaking Bad-length episode is affordable for a small production team. Getting sustained performances and continuity across an episode is still much harder than generating enough footage. When those are solved, the cost structure of television changes completely, and formats that exist only because production is expensive, like the one-minute microdrama, lose their reason to be micro. I think we’ll see the first genuinely good AI-made long-form drama within a few years. It will be good because of its writing and direction, the parts that were never about cost.
One more date worth noting: OpenAI’s Sora API shuts down on the 24th. The price of making video keeps falling. The price of making something worth watching hasn’t moved much, and I suspect it won’t.