The workstream
AI video costs ≈$88/video. We instrument every paid call, deploy one lever at a time, and measure each dollar saved.
Seedance 2.0 drives ~87% of known generation spend. Two views below: the live pipeline dashboard that measures the production run, and a worked example of the biggest lever — swapping the expensive model stack for a single self-hosted one.
~$88
Est. cost / video (n=23)
~37×
Cheaper on the video step (H3, worked example)
1 : 1
Every saving vs. measured baseline
Live instrumentation
Pipeline Cost & Flow Dashboard
The 9-step generation pipeline, instrumented end to end: per-step cost, time, tokens and retries, zoom-in call detail, tracked model alternatives, and a live-run status check.
9 stepsper-call costretry countslive status
Open the dashboard
Model-substitution lever
MiniMax Experiments
How much harder is a cheap self-hosted model than the expensive multi-model stack? A series of experiments rebuilding real production beats on MiniMax-H3, each scored for difficulty — starting with g7-cells-01 ($107.92 → ≈$1).
~37× cheaper video step≈2.7× more effortside-by-side frames
See the experiments
Evidence & documents
Plan & README
Phased workstream, primary sources, tooling.
Cost breakdown
Per-video estimates across n=23, spend concentration.
Strategy catalog
S1–S8 levers with savings estimates and risks.
MiniMax-H3 eval
Benchmark plan for H3 as a cheaper B-roll generator.
Savings methodology
Baseline, lever registry, per-change formulas.
Repository
Source, issues, and deploy notes.