One Head Chef, Many Helpers: Fewer AI Tokens
In September 2026 I started making my AI Safety shorts. Small videos, animation built from code. One episode used 3.2 million tokens. I sat there thinking, a video this small and a bill this big?
The mistake was mine. At first I did everything in Claude Code with the biggest model. Research, code, files, and the same checks again and again. Then when I started using agents, one single agent carried the whole old story with it into every new episode. In other words, the head chef was busy chopping onions. 😁
Think of the kitchen at a big wedding. The head chef decides the spices, tastes the food, and decides when the pot is ready. Chopping onions, peeling garlic, washing the pots, that is the helpers’ job. If the head chef spends the whole day on onions, the food is never ready on time.
I did the same with AI. My head chef is Opus 5.5, and Fable 5.1 when the job is really hard. It makes the plan, writes the script and the storyboard, and at the end checks everything itself. The helpers are Sonnet 5.5. They write code, build files, do research and run the renders.
Look at the price difference too. On the Anthropic API, writing 1 million tokens costs $50 on Fable 5.1, $20 on Opus 5.5 and $10 on Sonnet 5.5. So a helper does the work at half the price of Opus and one fifth the price of Fable. On a subscription it is the same story: the big model uses up your limit faster.
But to be honest, the real saving did not come only from the cheaper model. Moving the heavy work to the cheaper model is one reason. The second is that the head chef’s mind stays clear. It does not read every file itself, it only reads the helper’s short report. The third, maybe the biggest, is that now every episode gets a fresh helper who arrives with a playbook. Nobody re-reads the whole old history. And work that needs no thinking at all, like rendering the music or joining the video, runs as a plain command. Zero tokens.
The result: episodes 2, 3 and 4 took 1.7 to 3.2 million tokens each. Episode 5 took 446 thousand, episode 6 took 357 thousand and episode 7 took 322 thousand. Roughly 75 to 85 percent less. And the part I liked most, the fix rounds also dropped from 2 or 3 to just 1. Quality did not drop, because the decisions and the checking still sit with the head chef.
The same method built 100 shorts, the Teachers’ Day film, the gratitude video, 64 Daily Dhikr reels and 444 images, and a 20 second ad for a client. And this video too. Opus 5.5 wrote the script and the storyboard, two Sonnet 5.5 helpers built the code, animation and music at the same time, and the head chef checked every scene. Then a third helper fixed the problems the head chef found. The three helpers together did about 629 thousand tokens of work, while the head chef only had to read about 230 thousand tokens in the whole conversation. One fix round.
How you can do this
If you use Claude Code, this is an afternoon’s work.
First, keep your main session on Opus 5.5. For a really hard project, Fable 5.1.
Second, write one rule in your CLAUDE.md file: routine work like research, building files, code and repeated checks goes to Sonnet agents, and the main model only plans, reviews and decides.
Third, write a playbook for any job you repeat. One file that says how the job is done, where the files live and what a good result looks like. Then you don’t have to give the head chef long instructions every time.
Fourth, a fresh agent for every job. If you keep one agent running for months, it re-reads the whole old story every time, and that is where the tokens burn.
Fifth, if a step is just a command, keep it a command. Don’t turn it into an agent.
And last, never skip the review. The head chef must look at the helper’s work and send it back once with one clear instruction.
Don’t hand every small thing to an agent either. A one-line change the head chef should just make, because explaining it to an agent takes longer than doing it.
The head chef’s job is not to do everything. The head chef’s job is to make sure the work is right.
If anyone wants help with the setup, let me know and I’ll share my files. And if I got something wrong, corrections are always welcome.