After a few months of running generation through an agent, the surprising thing wasn't the cost per call. It was how much of the spend went to work I didn't need to do.
Almost every fix below is about ordering and verification, not about paying less per call.
1. Iterate on the cheap tier. Always.
The largest available saving, and the most commonly skipped.
Most model families ship a fast or lite variant. They exist for exactly this: when you're deciding composition, framing or timing, you're making a decision, not producing a deliverable.
The number of drafts you discard is much larger than the number of finals you keep. Optimising the discarded ones is where the leverage is.
The failure mode is exploring on the expensive model because it produces nicer images. It does — and you're paying production prices for drafts you're about to throw away.
2. Front-load the local tools
A significant part of any video pipeline needs no model at all: transcription, scene detection, silence cutting, word-boundary alignment, and every validator.
Two consequences. First, they're free, so run them liberally. Second, and bigger:
Reduce before the model sees it. A 90-minute transcript folded to phrase level is a fraction of the input. The model doesn't need the full haystack — and producing the needle is usually a local operation.
3. Pilot every batch
Run three. Inspect. Then run the rest.
Batch work is the one place where cost is genuinely predictable — one input, one output, roughly linear. So a three-item pilot gives a real projection: measure actual consumption, multiply, add 20%.
But the real value is qualitative. Whatever is wrong with your prompt is wrong in every item. Finding out on three costs three.
4. Calibrate before generating narration
Speech rate is not a number you can look up. It varies by voice, language, punctuation, and the specific text.
tts_pacing_calibrate # chars-per-second from one sample, project total runtime
Measure on one sample, project the runtime, adjust script or speed, then generate. A full narration pass that doesn't fit the edit is a complete write-off — and entirely avoidable.
Worth being honest about the failure mode here: there's a temptation to solve a script problem with a speed parameter. It doesn't work. It converts "too long" into "sounds rushed."
5. Validate every render, not just the final
This one saves money in a less obvious way.
The validators cost nothing, so the question isn't whether to run them but when. Run them on every render, including intermediates — because:
A defect caught at the intermediate stage costs one regeneration. The same defect caught after assembly costs the assembly too.
6. Reference images instead of adjectives
Describing a look in words takes several attempts to converge. Supplying one reference image often gets there in one.
Every failed attempt is a full-price generation. Three rounds of warmer, softer, less contrast costs three generations and still lands somewhere approximate.
One good reference image replaces a paragraph of adjectives, and it's more reliable.
7. Keep a manifest
Not obviously a cost measure. It is one.
input | prompt | model | generateId | output path | status
Without this you re-derive that information by hand, and when a run dies partway you restart from the beginning instead of from where it stopped.
A manifest is a checkpoint, and checkpoints are what make long jobs survivable.
The one I'd leave off the list
I'd stop short of recommending you optimise the model choice itself beyond the fast/full split.
Chasing per-model price differences produces small savings, a lot of context-switching, and pushes you toward models you don't know well.
Knowing one model deeply is worth more than a marginal rate — you get what you wanted in fewer attempts, and attempts are the actual cost.
The pattern underneath
Look at the list again. Almost none of it is "spend less per call."
It's: decide cheaply, verify early, checkpoint often, and don't re-derive what you already knew.
Which is the same list you'd write for any expensive batch process. The models are new; the discipline isn't.
curl -s https://files.dlazy.com/cdn/cli | bash
dlazy -h # tool list
dlazy <tool> -h # flags — read this before the first paid call
That second command is the cheapest habit on this page. Flag sets differ more than you'd expect between tools, and a wrong guess costs a generation.
Top comments (0)