Methodology

Model ≠ platform ≠ plan

A recommendation is a model (what generates), a platform (the product and workflow controls it is used through), and a plan (what you pay for and how much capacity it includes). Veytora AI keeps the three separate and never collapses them into one “tool score”. An AI model may perform differently depending on the platform, workflow controls, and plan through which it is accessed. Cheapest is not always best value.

Two-stage decision engine

Stage A — Fit Gate

Candidates that fail a verified hard requirement are removed before any ranking, and the reason is stored: wrong input mode for your source material (image-to-video only when no starting image is confirmed), a verified maximum video length below your output, a generation length that cannot reach your scene, a strict budget, or a non-negotiable that requires a capability the platform has not verified. Gates use only factual data. A workflow can never omit a required layer: if every candidate for a required layer is rejected, no recommendation is shown at all and the rejected candidates are listed under “Why not”.

Capabilities are evaluated as a set

Every capability you select is checked against the whole stack. One subscription may cover several of them — OpusClip Starter and above cover clipping, captions, and shorts editing — so a second tool is only added when it supplies a required capability the first does not. Where support is not verified for any tool, the stack is rejected with “not enough verified evidence”; a selected capability is never silently ignored. Stacks that cost more, or add subscriptions, without adding a required capability, capacity, or evidence are pruned before ranking.

One base plan per provider

For providers whose subscription credits are shared across several production layers, the engine selects one base subscription plan and aggregates all relevant usage into that plan’s shared allowance. It does not recommend stacking multiple mutually exclusive base plans from the same provider simply to increase included capacity. Where a provider verifiably sells extra credits, that shortfall is priced separately as an estimate.

Capacity status

• Verified fit — every required layer is present, every selected capability is covered, and every capacity comparison is based on verified, non-approximate data that covers the modelled workload.
• Partly verified — the workflow is complete and nothing verified is short, but at least one layer’s capacity is unknown or provider-approximate (“≈40 videos”). Shown as the known subscription base cost, never as “fits your workload”.
• Insufficient — a verified pool is below the workload. Never the primary recommendation; shown only as the “closest capacity option” with its exact shortfall.

Annual pools are compared as monthly workload × 12 against the annual amount and displayed annually (3,600 credits/year, with the monthly equivalent labelled as such). Rollover is shown as flexibility and is never counted as baseline capacity. Where a provider verifiably sells extra credits (Runway: 1,000-credit minimum at $0.01/credit), a shortfall is labelled “requires purchased credits” with an estimate that is explicitly not exact; strict-budget checks stay unavailable in that case.

Stage B — Ranking

Eligible, complete stacks are ordered: verified fit → partly verified → insufficient; then fewest critical requirements without verified evidence; then verified subjective production-fit evidence where it actually exists; then lower known cost; then fewer subscriptions; then a deterministic tie-break. Because no independent quality evidence is recorded today, Best Value currently behaves as the lowest-cost complete, qualified, verified solution — not an editorial quality score divided by price. Editorial workflow-fit numbers in the data files are legacy display values and do not affect ranking. Cost is never normalised against the most expensive candidate, so adding an irrelevant expensive plan cannot change anyone else’s position.

Job-specific priorities

Each creator job (long-form/faceless, Shorts from long-form, AI-native Shorts, image-to-video scenes) has its own weight preset for subjective production-fit dimensions. These weights are product heuristics, not scientific facts, and only take effect once verified evidence exists for those dimensions.

Profiles

• Best Value — strongest verified fit relative to cost; with today’s evidence, the lowest-cost complete verified solution.
• Lowest Cost — the lowest-cost qualified option: complete, every selected capability verified, and capacity not known to be insufficient.
• Quality First — shown only when independent or internal quality evidence exists; otherwise listed under “Additional comparisons” with “Not enough comparable quality evidence yet”.
• Fastest Workflow — shown only when verified speed or throughput evidence exists. No generation-speed numbers are invented.

Evidence and confidence

Every dynamic fact carries a source type, source, and verification date. Official pricing and specifications can give high confidence for factual fields. An official marketing claim is never treated as quality evidence and never produces a numeric quality score. Unknown quality data stays unknown.

Retry-adjusted usable cost

Nominal cost is what the provider charges per generation. Effective cost per usable output can only be calculated once a retry assumption exists, and Veytora AI does not invent one: when you leave it blank the result is labelled nominal and retry-adjusted cost is marked unavailable.

Prices, billing, and savings language

Annual plans are shown as the amount billed upfront plus a monthly equivalent (“$174/year billed annually · $14.50/month equivalent”); the equivalent is a display convenience, never monthly billing, and an annual commitment is always surfaced as a trade-off. In Cut My AI Spend, differences are “potential” and based on known pricing; they are only exact when both prices are verified, capacity is verified (not approximate), and the billing period is comparable. Prices you enter are labelled user-provided.

Switch and downgrade conservatism

A downgrade is offered only when the lower plan retains every verified plan-level feature and verified capacity covers your usage; approximate capacity yields “downgrade worth reviewing”. Another provider is only an “alternative worth reviewing” — never a directive — when it covers every verified capability and usage and clears max($10, 15% of current cost). A subscription is marked redundant only when another current plan covers all its capabilities and has verified capacity for the combined usage — and that covering plan must itself remain. When two plans could each cover the other, exactly one deterministic survivor is kept (verified capacity, then lower cost, then a stable tie-break); circular dependencies fall back to “review”. The “after confirmed changes” total includes only confirmed actions (keep, verified downgrade, safe removal). Alternatives worth reviewing — other-provider switches, approximate or annual-billed downgrades — are shown individually with their own potential difference and are never summed into a stack total or shared as savings. Smaller differences return “keep”. Affiliate economics never enter any of this.

What a recommendation is built from

1. Your workload — videos per month, finished minutes, AI-video coverage, scene length, images, and voice minutes.
2. Verified provider limits — plan prices and included credits or minutes, each with a last-verified date.
3. Retry assumptions — attempts per usable clip and per usable image. Every attempt is billed, so this multiplies everything.
4. Duration padding — when a model only generates fixed lengths, a 7-second scene becomes a 10-second generation. You pay for the 3 seconds you cut.
5. Subscription capacity — required credits ÷ included credits. Over 100% means the plan does not fit.
6. Editorial quality and workflow scores — judgments, not measurements. See below.

Workload formulas

With V videos, M finished minutes, P AI-video coverage, S average scene seconds, A attempts per usable clip, I images per video, and T voice minutes per video:

• AI-video seconds = V × M × 60 × P
• Usable clips = ⌈AI-video seconds ÷ S⌉ (or V × your shot count if you provide one)
• Estimated scene attempts = ⌈usable clips × A⌉; retry scene attempts = scene attempts − usable clips
• Generation calls = scene attempts × segments per attempt (a 13-second scene on a 2–10 s model is 10 s + 3 s: two calls per attempt)
• Required credits = scene attempts × credits for one complete routed attempt (all segments)
• Generated seconds = scene attempts × routed seconds per attempt; duration padding = scene attempts × (routed seconds − scene seconds)
• Image generations = ⌈V × I × image attempts⌉; voice minutes = V × T

Rounding rule

A scene attempt is an indivisible billing event, so the attempt count is rounded up first and every downstream figure — credits, generated seconds, padding, generation calls — is derived from that integer. A scene attempt and a generation call are different things: for a multi-segment route (10 s + 3 s for a 13-second scene) one attempt is two provider calls. Veytora AI shows both and never mixes a fractional expected attempt count into one number and a rounded count into another.

Shared subscriptions and pools

A single subscription can serve more than one layer — a Runway plan’s credits are one pool drawn on by both video and image generation. Veytora AI enumerates every combination of verified plans, groups the chosen plans into subscriptions, adds up all the usage each shared pool receives, and only then tests fit. A plan’s price is counted once per stack no matter how many layers use it, and a shared pool is never split into per-layer copies. When no combination fits, the stack with the smallest normalized shortfall (required − included, divided by required) is shown and labelled “closest fixed-credit option”; the three profiles may converge on the same answer.

Native resolution vs delivery resolution

A model’s native output resolution is recorded separately from the resolution you choose to deliver. When they differ, the model is still recommended, but the upscale or post-processing step is flagged and its cost — which Veytora AI has not verified — is excluded, so the estimate is marked partial.

Legacy plans and purchased credits

Plans no longer sold to new subscribers (Runway’s Unlimited plan, legacy after June 1, 2026) are kept in the data only as legacy records and are never recommended. Fit is always tested against fixed included credits for a single month, with no rollover assumed. Where a provider verifiably sells extra credits, the shortfall is priced separately at the verified rate and minimum purchase, and labelled an estimate.

When the estimate is incomplete

A stack’s dollar figure is labelled a complete estimate only when every required layer has a verified recommendation and every capacity-limited plan covers the workload. Otherwise it is shown as the known subscription base cost, marked partial or insufficient, and never compared with your budget — when required top-up pricing, overage pricing, or provider capacity is unverified, Veytora AI does not treat the total as final. Where top-up pricing is verified (Runway), the shortfall is priced separately as an estimate.

Capacity labels

• 0–25% — Large unused capacity
• 25–60% — Comfortable headroom
• 60–90% — Good fit
• 90–100% — Tight fit
• Over 100% — Insufficient included capacity

These labels are advisory, not absolute claims. A plan that does not fit is never recommended over one that does, and Veytora AI never invents overage pricing to make it look affordable.

Stack scoring weights

Each layer’s candidates receive 0–100 sub-scores, combined with these weights:

Editorial scores

Editorial quality and workflow-fit numbers still exist in the data files with a review date. They are opinions, not benchmarks. In the Advanced Planner they remain a labelled, secondary factor; in Check they are display-only and are never used to rank recommendations.

What never enters the ranking

Affiliate commission percentage, affiliate approval status, referral payout, and cookie duration are never read by the recommendation engine. A test in the codebase changes every commission value and asserts the recommendation order is unchanged. Affiliate status changes only which link a button points to, and the disclosure shown next to it.

Affiliate relationships never affect Veytora AI rankings.

Prices change. Models change.

Every plan and model carries a last-verified date. If it is older than 14 days, a warning appears. When a value is missing — a generation cost, a regional price — the calculator says so rather than estimating.

• Runway — Last verified: September 10, 2026
• ElevenLabs — Last verified: September 10, 2026
• Descript — Last verified: September 10, 2026
• Google Flow — Last verified: September 10, 2026
• Magnific — Last verified: September 10, 2026
• CapCut — Last verified: September 10, 2026
• OpusClip — Last verified: September 14, 2026
• Submagic — Last verified: September 14, 2026

Seed data was verified on 2026-09-10. Read the affiliate disclosure for how links are handled.