7.5 The Cost Architecture of the Pipeline
Chapter 0.4 gave the break-even calculation. This chapter is the structure that makes the number predictable: what the tiers actually buy, where money leaks, and the two decisions that determine whether the pipeline costs tens or hundreds of dollars a month.
[AS OF 2026-10] Prices are point-in-time and move constantly. Verify before budgeting.
Three tiers
A practical structure, built from current per-unit pricing rather than from any vendor's package.
| Tier | Monthly | Composition | Suitable for |
|---|---|---|---|
| Minimum | ~$36 | One image subscription, a voice tier, no video | Stills and voiceover; one persona; learning the pipeline |
| Working | ~$102 | Image subscription or API credit, voice mid-tier, a video budget, lip-sync for occasional talking posts | A posting persona with some video; the realistic operating tier |
| Volume | ~$212 | Higher generation capacity, video-heavy, dedicated sync minutes, storage | Daily posting, multiple formats, brand work with variation requirements |
What each tier actually changes. Not quality — all three can produce publishable output. The difference is how much iteration you can afford, and iteration is what determines the rejection rate, which is what determines publishing volume.
The tier you need is a function of your format mix, not your ambition. An account publishing stills with voiceover runs comfortably in Minimum. The same account adding one synced A synced speaking clip; the format that moves a pipeline from Minimum to Working. a week moves to Working, because sync minutes plus retries dominate the marginal cost (6.6).
Where the money actually goes
A working-month breakdown, illustration only and heavily dependent on volume:
| Line | Share | Notes |
|---|---|---|
| Image generation | small | Cheapest line per unit; volume offsets it |
| Video generation | largest | Per-second pricing × rejection rate |
| Voice | smallest | Never the problem (6.5) |
| Lip-sync | moderate | Per-minute, multiplied by languages |
| Storage and backup | small | Grows quietly; a fixed cost you should measure |
| Subscriptions | fixed | Paid whether used or not |
| Human time | largest total | Not cash, but the binding constraint |
The line that surprises people: storage. Generated assets accumulate, and 4K video masters are large. A year of daily posting can be hundreds of gigabytes, and the cloud copy doubles it. This is a small number that becomes a real one, and the fix is a policy rather than a tool: keep masters for published work, keep references and adapters forever, and discard raw takes systematically.
Three leaks
Credit expiry. Prepaid credits that expire are a total loss on the unused balance, and chapter 1.5's fourth failure pattern. The mechanism is subtle: you buy capacity for a burst, the burst is smaller than planned, and the remainder evaporates. The policy: buy capacity slightly ahead of demonstrated need, never in bulk. If a vendor's discount requires a bulk purchase, calculate whether the discount exceeds the expected expiry loss at your real consumption rate — usually it does not.
No rollover on subscriptions. A monthly subscription's quota does not carry over. If you consume half of it, you paid double for what you used. The policy: measure consumption for two months before choosing a tier, and step down if utilization is under roughly a third. Chapter 5.6's survival rate is the input to that measurement.
Mid-month exhaustion. Running dry on the twentieth is not bad luck; it is a signal. Three readings, in order of likelihood:
The model is wrong for the job. A premium tier for background clips in a montage is money burned; chapter 6.2's tier-follows-role rule exists for this.
The rejection rate rose. Something in the reference layer or the The maintained set of forty reusable prompt variants, refilled in the weekly review slot. regressed, and you are generating more to get the same output. The rejection rate is the diagnostic, which is why it is tracked.
The volume genuinely grew. The good version, and it means raising the tier deliberately rather than exhausting credit accidentally.
A pipeline that exhausts its budget before the month ends is giving you information. The three possible readings have three different fixes, and only one of them is "spend more."
The two-vendor rule for video
Chapter 6.1 described a frontier video product being discontinued during the research period for this book — the app first, the API later. [AS OF 2026-10] The operational lesson is a rule:
Maintain working knowledge of two video providers. Not two subscriptions; two known paths. When one changes terms or disappears, the migration is a week of testing rather than a month of lost output.
What "working knowledge" means concretely: you have generated test clips on the second Under the AI Act, whoever develops an AI system or places it on the market under its own name., your keyframes and prompts are provider-neutral (still-first workflow, 6.1), and your asset library does not depend on a provider-specific format.
This is also why the still-first workflow is a cost decision. Validated keyframes are portable across every video model. A library of vendor-specific generated clips is not.
When self-hosting beats API
The arithmetic, stated plainly.
Self-hosting wins when: you generate a lot, consistently. A rented GPU at roughly $0.50–$2.00 per hour (3.3) replaces per-unit pricing with per-hour pricing, and the crossover arrives quickly for image generation at volume. At a few hundred images a month the crossover has already happened for images; for video, the open-weights route removes per-A CLIP-style text encoder that responds best to comma-separated fragments; used by older model families. cost entirely and is the decisive factor at high volume.
Self-hosting loses when: your volume is variable, you do not want to operate a machine, or your time is worth more than the difference. The setup and maintenance burden is real — drivers, updates, a machine that must be on, and the failures in 7.2's table which are yours to fix.
The break-even shape, roughly:
| Monthly generation | Likely cheaper route |
|---|---|
| Under a few hundred images, light video | Subscription |
| Hundreds of images, regular video | Either; measure |
| Thousands of images, daily video | Rented GPU |
| Any volume with no tolerance for ops work | Hosted, permanently |
And one warning about the self-hosted route's hidden cost: the hours. A weekend of setup is a weekend, and it recurs when something breaks. If you are not the kind of operator who enjoys that, the hosted tier is cheaper than it looks — because the alternative is a pipeline that stops when you stop maintaining it.
Cost per unit of output
The metric that makes decisions comparable. Two numbers, both defined in 5.6:
Generation spend divided by frames passing QA; the metric that makes tier changes comparable. = total generation spend ÷ frames passing QA.
Monthly pipeline spend divided by posts published; below roughly $6 is healthy, above $15 means step down. = total monthly pipeline spend ÷ posts published.
| Scenario | Images generated | Survival | Spend | Per usable | Posts | Per post |
|---|---|---|---|---|---|---|
| Subscription, established pipeline | 150 | 1 in 3 | ~$36 allocation | ~$0.24 | 4 | ~$9 + time |
| Subscription, regression | 400 | 1 in 8 | ~$36 allocation | ~$0.48 | 4 | ~$9 + time |
| Per-image API | 150 | 1 in 3 | $54 | $1.08 | 4 | $13.50 |
What the table is for. Two things. It shows that a regression in the rejection rate doubles your effective cost without changing a single price — which is the argument for fixing the reference layer instead of buying more capacity. And it shows that the difference between routes is much smaller per published post than the headline per-image price suggests, because the subscription route's unused capacity is also a cost.
Track these two numbers monthly. They are the only figures that let you compare a tier change, a model change, or a self-hosting decision on the same basis.
Where this goes wrong
Buying volume for a burst that does not arrive. Expiring credits, total loss.
Choosing a A recurring access level priced for entry, mid, or high intent. before measuring consumption. Paying double for what you use.
A premium model for low-value shots. Tier must follow the shot's role.
Reading mid-month exhaustion as "spend more." Usually it is the model or the rejection rate.
One video vendor. One discontinuation from a lost quarter.
Self-hosting without counting the hours. The hidden cost is the one that recurs.
No cost-per-usable-frame metric. Every comparison becomes a guess.
The gate
- Do you know which of the three tiers you are actually in, and what your format mix requires?
- Have you measured your subscription utilization for two months?
- Do you track rejections, and would you notice a regression in them?
- Do you have working knowledge of two video providers?
- Have you done the self-hosting arithmetic including your own hours?
- Do you know your cost per usable frame and per published post?
What you leave with
Three tiers defined by iteration capacity rather than quality, a spending breakdown with the storage line people forget, three leaks with a policy for each, the three readings of mid-month exhaustion, the two-vendor rule with the asset-portability argument behind it, a self-hosting break-even shape including the hidden cost of hours, and the two metrics — cost per usable frame and per published post — that make every future decision comparable.