Cost is the easy question — swapping the video model is ~37× cheaper on that step (about $36 off the $108 all-in; the shared planning + keyframes are the same either way). The real question for adopting MiniMax-H3 is effort: what work moves onto you when you drop the autonomous multi-model pipeline for one self-hosted image-to-video model? Each experiment below rebuilds a real production beat and scores the difficulty, axis by axis.
Aggregated across the experiments so far. The scorecard is the shared rubric every experiment fills in; the number moves as more beats are tested.
The honest answer: MiniMax-H3 is meaningfully harder to operate than the expensive stack — but almost all of the extra difficulty is one-time and structural, not recurring. Three things drive it: labels move to a separate overlay stage, motion must be faked in post, and you run your own GPU. What you don't lose is the look — i2v conditioning on the existing keyframes reproduces it. So the trade is real work and lower autonomy in exchange for a ~37× cut on the video-generation step (about $36 off the $108 all-in — see the provenance breakdown). Once the playbook, prompt templates and a warm GPU image exist, the per-video gap narrows to prompt-writing plus label authoring.
| Dimension | Expensive stack | MiniMax-H3 | Δ | What the difference is |
|---|---|---|---|---|
| Prompt authoring | Expensive stack | MiniMax-H3 | +2 | Expensive stack assembles a structured contract across LLM stages automatically. H3 wants one hand-written paragraph per scene — you must know the style→subject→hold→no-text scaffold. |
| First-frame consistency | Expensive stack | MiniMax-H3 | 0 | Original locks identity with an asset-bible master PNG + sha, passed to reference-to-video. H3 conditions on the same keyframe — equally easy when keyframes already exist. |
| Iterating to an acceptable take | Expensive stack | MiniMax-H3 | +3 | Original auto-retries in the cloud (hands-off, but you pay per attempt — 144 here). H3 iteration is manual and you must diagnose failure modes (turbo bubbles, drift, clip length) — heavy the first time, light once learned. |
| On-screen text & labels | Expensive stack | MiniMax-H3 | +3 | Original requests labels via a native_labels field and the model/label stage renders them. H3 garbles any text, so labels move to a separate HyperFrames overlay stage you author by hand — the single biggest added burden. |
| Motion & camera | Expensive stack | MiniMax-H3 | +2 | Original takes rich camera language directly. H3 must be told to hold still, then motion is faked with hold-to-slot / Ken-Burns in post — extra editing and a lower motion ceiling. |
| Audio | Expensive stack | MiniMax-H3 | 0 | Original runs a dedicated TTS/music stage. H3 bakes a soft ambient bed you high-pass and mix under reused VO — comparable effort, different shape. |
| Infra & ops | Expensive stack | MiniMax-H3 | +3 | Original is pure API calls — a key and a budget. H3 means standing up a GPU box, ComfyUI, ~60 GB of weights and a driver script before a single frame renders. |
| Autonomy (idea → video) | Expensive stack | MiniMax-H3 | +2 | Original goes from a topic to a finished video unattended. H3 needs keyframes and narration to already exist and a supervisor tuning each beat — in the g7-cells rebuild that supervisor was Claude Opus 4.8, doing it by hand. |
Dots = human/ops friction (1 trivial → 5 heavy). Δ is how much harder H3 is on that axis (red = harder). Cost runs the opposite direction — but honestly: the expensive stack is ~37× more expensive on the video-generation step H3 swaps (about $36 / 33% off the $108 all-in; see provenance).
The live research log — claims we're testing, each with its status and the evidence so far. New findings get appended here and linked to the experiment that produced them.
An orchestrator layer with an inexpensive model + MiniMax-H3 is overall more efficient than the expensive multi-model stack.
Evidence. muse-glimmer-30b drove render→QC→judge on SC3 and autonomously chose the correct fix (full_steps — drop turbo LoRA, 20-step) that Claude Opus 4.8 found by hand supervising the g7-cells rebuild — for a fraction of a cent. Slow (~8s/turn) but correct. · orchestrator experiment →
Cheap vision models can reliably detect H3's failure modes — especially on-screen text, its #1 defect.
Evidence. All three candidates flagged baked captions (glimmer read the exact caption text); qwen/inkling passed a clean frame with 0.92–0.99 confidence. Deterministic OCR false-positives on flat art, so the vision judge is the reliable text detector. · the eval →
i2v conditioning on the original keyframes reproduces the look — the keyframe replaces the whole asset-bible.
Evidence. g7-cells-01: 34/34 keyframes reused byte-identical, visual parity across all 17 beats; the hero cell stays on-model without any entity-bible machinery. · g7-cells-01 provenance →
Swapping only the video model saves ~1/3 of all-in cost (~$36), not ~108×.
Evidence. Authoritative decomposition: shared upstream $70.97 (planning + keyframes) + video $36.95 = $107.92, vs $70.97 + ~$1 ≈ $72. The video-generation step alone is ~37×; the ~108× only holds when re-rendering off existing keyframes. · cost decomposition →
A near-frontier driver (Opus / GLM-5 / Kimi-K3) converges in fewer iterations than a cheap one — worth its higher per-call cost only on hard beats.
Evidence. To test in the driver bake-off (exp 05): same beats, driver swapped across models, scored on iterations-to-acceptable and final on-model rate.
Even from zero (no keyframes), H3 + a cheap image model + TTS beats the full autonomous stack on cost.
Evidence. The from-zero experiment removes H3's reused-keyframe advantage to measure the true floor against the $107.92 all-in run.
Two with results, two planned — chosen to probe where the effort gap is widest (labels, motion), whether a cheap model can run the loop, and where cost should nearly vanish (from-zero).
Flat-vector Grade-7 biology explainer with a recurring hero cell. 17 beats reproduced by i2v on the original keyframes; labels moved to HyperFrames.
Can a cheap model do the supervision Claude Opus 4.8 did by hand? muse-glimmer-30b drove render→QC→judge and chose the correct fix autonomously; qwen/inkling judge the frames.
Same beats, swap the driver across glimmer / kimi-k3 / glm-5p2 / deepseek / opus. Score iterations-to-acceptable and final on-model rate — does frontier converge faster enough to justify its cost?
Remove the biggest H3 advantage: start from just a topic. Measures the true floor — one image model for keyframes + TTS + H3 — against the full autonomous run.
Every experiment rebuilds a real beat (or whole video) three ways to hold the output constant: same keyframes, same narration, same timeline. We then record, per beat, the exact machinery each side ran — models, retries, entity locks, QC gates for the original; workflow, steps, seed, prompt for H3 — and score eight axes of human/ops friction from 1 (trivial) to 5 (heavy). Cost comes from the pipeline's own billed ledger, not estimates.
“Harder” is deliberately about effort, not quality or price — those are reported separately so a reader can weigh all three.