Draft Tiers Are the Cheat Code for AI Video Budgets

AI video cost control usually starts in the wrong place: teams shop for a cheaper model when the cheaper option is already sitting inside the model they picked. FLUX 3 Video, MiniMax H3 and Veo 3.1 all publish a low-cost exploration tier alongside their full-quality render. Used deliberately, that tier is the single largest lever on a creative budget — bigger than switching vendors, and available today without renegotiating anything.

The reason it goes unused is that most teams treat generation as a single act. Prompt, render at final quality, look at it, dislike it, render again at final quality. Every rejected attempt is billed at the price of a shot you intend to ship. On a campaign where you explore several directions per shot, the majority of the bill is paid for footage nobody will ever see.

This guide takes the opposite approach. It reads the cheap tier off each vendor’s own pricing page, works out what the discount actually is, and then asks the question nobody else is asking: does the vendor document a mechanism that carries your approved cheap draft into the expensive finish? On two of the three platforms it does. On the third — the one with the steepest discount — it does not, and that changes the recommendation completely.

Start with the raw rates, all read from vendor pricing pages at the time of writing. Black Forest Labs prices FLUX 3 Video text-to-video and image-to-video at $0.17 per second for an HD render, $0.29 per second for FHD, and $0.06 per second for a draft. Video continuation sits on a separate, higher table: $0.43 per second HD, $0.54 FHD, $0.12 draft. Those two tables are not interchangeable, and mixing them is the fastest way to under-budget a project that leans on continuation.

MiniMax prices H3 at $0.13 per second for 2K output and $0.08 per second for 768P, with a separate regeneration path at $0.05 per second. Google prices Veo 3.1 in three tiers: Standard at $0.40 per second for 720p and 1080p and $0.60 for 4K, Fast at $0.10 / $0.12 / $0.30, and Lite at $0.05 per second at 720p and $0.08 at 1080p, with no 4K output on Lite at all.

Convert each cheap tier into a percentage of its own full tier and the picture stops being a price list and starts being a strategy. The division below is ours, performed on the vendor-published rates above so it can be checked line by line.

Read that chart carefully, because the intuitive conclusion is the wrong one. Veo 3.1 Lite offers the deepest discount and MiniMax 768P the shallowest, so a naive cost optimiser would explore on Veo Lite. The rest of this guide argues the opposite — that the depth of the discount matters far less than whether the vendor has documented a way to turn the cheap thing you approved into the expensive thing you ship. That is the axis nobody publishes, and it inverts the ranking.

If you want the provider-by-provider price list rather than the workflow, we maintain the full multi-provider pricing table separately. This post assumes you have already chosen a model and wants to change how you spend on it.

BFL’s FLUX 3 Video documentation describes a two-call pattern. Setting draft: true on a request returns a fast preview instead of a full render, and hands back a draft_cache bundle alongside the result. When you have picked the preview you want, you submit a second request in mode: “draft_enhance”, pass that bundle back as draft_cache, and the original generation is reproduced at full quality.

The docs are explicit about why this is different from simply re-prompting: each fresh submit is its own generation, so a direct full render can interpret the shot differently from the draft you liked. That single sentence is the reason a draft-first discipline is reliable on FLUX 3 rather than merely cheap. You are not hoping the expensive render lands in the same place as the cheap one. You are asking the API to replay it.

Every mode outputs 24 frames per second at HD or FHD, with FHD documented as 1920 × 1088 for 16:9, and aspect ratios from 21:9 through 9:16. Drafts render at HD. All three modes run on the same flux-3-video endpoint, and requests are asynchronous — you submit, receive an id, and poll.

One caveat has to sit next to any recommendation to build on this: BFL’s own documentation states that FLUX 3 is a preview model, with video editing and Omni Reference described as coming later. A preview label is not a reason to avoid it, but it is a reason to keep the integration thin and to avoid hard-coding assumptions about request shapes that may still move. For the platform’s full feature set we covered FLUX 3 Video’s general-availability release separately; the general-availability date carried in that coverage comes from BFL’s launch blog rather than the pricing docs used here.

MiniMax structures the same idea differently. Rather than a draft flag on the generation request, it exposes a distinct MiniMax-H3-Regeneration path whose entire job is to take a previously produced 768P video and re-render it at 2K, billed per second of the regenerated output at $0.05.

The arithmetic is less flattering than FLUX 3’s at first glance. At $0.08 per second, 768P is 61.5% of the $0.13 2K rate by our own division — a discount, but not a dramatic one. The ladder still wins, and for a reason worth stating precisely: the regeneration leg is cheaper than either generation tier. You pay $0.08 per second only for the exploration, and the promotion of the winner costs less per second than generating it at 768P would have.

That last card is the caveat most secondary coverage omits. MiniMax’s own page states that the input materials used in the original 768P generation task are billed again on regeneration — audio free, images free for the first five and $0.025 each beyond that, video at $0.05 per second of the original input duration. So the clean $0.08 per second plus $0.05 per second ladder is exact only for plain text-to-video work, or image-to-video work using five or fewer reference images. Past five, add the re-billed images to the promotion leg or the model will read low.

One sourcing note: every rate above is MiniMax’s own direct pay-as-you-go price. If you buy H3 through a third-party gateway instead, price the work from that gateway’s rate card rather than this one. We covered MiniMax H3’s launch specifications when the model shipped.

Google’s Veo 3.1 lineup is where the cost logic and the workflow logic come apart. Three tiers ship as three separate preview models — veo-3.1-generate-preview, veo-3.1-fast-generate-preview and veo-3.1-lite-generate-preview — priced at $0.40, $0.10 and $0.05 per second respectively at 720p on the paid tier. Lite is the cheapest per-second video-with-audio rate on the board.

The problem is structural rather than commercial. Google’s pricing documentation prices three models; it does not describe a mechanism for carrying a composition approved on one tier into a render on another. There is no published equivalent of BFL’s draft_cache bundle or MiniMax’s regeneration path. A team that explores on Lite and finishes on Standard is not finishing anything — it is generating a second, unrelated video from a different model and hoping it resembles the first.

None of this makes Veo 3.1 a poor choice. Standard at $0.40 per second is a fully-featured render, Fast at $0.10 is a genuinely useful middle rung, and Lite is a reasonable way to test whether a prompt direction is worth pursuing at all. What it is not is the first rung of a locked ladder. Treat the three tiers as three independent options selected per shot, and the pricing is excellent. Treat them as draft and final, and you have bought a discount on a promise nobody made.

Two housekeeping notes on this family. All three Veo 3.1 IDs are still preview models, which Google flags as subject to change with more restrictive rate limits. And Veo 3 and Veo 2 both carried a shutdown date that has already passed — treat them as gone rather than as deprecating options, and do not plan around them.

Here is the comparison in one place. Every per-second figure is read from the vendor’s own pricing page; every ratio in the discount column is our division of those two figures, shown so it can be re-checked. The final column is the one that does not exist anywhere else — each vendor documents its own ratio, none of them documents whether a competitor guarantees the finish.

The Fast row is worth a second look. At 25.0% of Standard — $0.10 divided by $0.40 — Veo Fast is a deeper discount than MiniMax’s 768P tier and deeper than either of FLUX 3’s draft ratios, and it carries exactly the same absence of a lock-in mechanism as Lite does. If you are working on Veo, the honest framing is that you are choosing a quality-cost point per shot, not building a pipeline. That is a perfectly good way to work. It is simply a different one from the ladder this post is about.

Abstract ratios do not survive a budget meeting, so here is a concrete campaign. Six shots, five seconds each, thirty seconds of finished runtime. The team explores three variants per shot before locking one winner.

That three-variant figure is our own illustrative assumption, chosen because it is round and defensible — it is not an industry standard, not a vendor recommendation, and not drawn from any study. Your real number is whatever your creative director actually needs to see, and the model below is built so you can substitute it. Every cell is the product of the stated formula and the vendor rates in the previous section.

Three things fall out of that table. First, the FLUX 3 route is the cheapest guaranteed path to a finished HD second in this model at $0.35, against $0.51 if you render every attempt at full quality. Second, MiniMax’s ladder produces a 2K finish at $0.29 per finished second, which is lower than FLUX 3’s HD figure — different output specifications, so read it as a cost data point rather than a like-for-like quality comparison.

Third, and most instructive: the Veo row shows the largest headline saving on the board and is the one we would not run. A 54.2% reduction that arrives with no mechanism guaranteeing the Standard render reproduces what was approved on Lite is not a saving, it is a deferred risk. If the Standard finish misses, you re-render at $0.40 per second until it lands, and the saving evaporates somewhere around the second attempt.

The MiniMax figures above assume plain text-to-video with no reference images, or five or fewer. If your winning shot used more than five reference images, the regeneration re-bills each additional image at $0.025, and those cents belong in the finish leg.

This is the sensitivity that decides whether the discipline is worth the extra API call. Solve the break-even algebraically and the FLUX 3 draft path costs less than direct full rendering from the second variant onward: at one variant per shot, a draft plus an enhance ($1.15 per shot) costs more than a single full render ($0.85). On MiniMax the two approaches tie exactly at one variant — $0.65 per shot either way — and the ladder wins from two variants up.

The practical reading is that draft tiers reward exploration and punish certainty. If a shot is fully specified and you expect to nail it in one pass, render it directly. The moment you expect to look at alternatives, the ladder is the cheaper instrument, and the gap widens the more you look. Teams that already budget generation spend against a cost ceiling can express this as a rule rather than a preference: exploration passes draw from the cheap tier, finishes draw from the expensive one, and the ratio between them is a number you can watch.

A cost framework that only works is not a framework, it is a sales pitch. Four boundaries matter here, and the first one rules out an entire model.

The second boundary is mode confusion. FLUX 3 Video’s continuation table is a different table, and its draft ratio (27.9% of HD by our division) is not the roughly one-third ratio BFL quotes for text-to-video. Any spreadsheet that applies one rate card to both modes will under-budget continuation-heavy work by a wide margin.

The third is reference-material re-billing on MiniMax, covered above: the two-stage arithmetic holds only for plain text-to-video or for image-to-video using five or fewer reference images. The fourth is the preview status of the platforms themselves — FLUX 3 is documented as a preview model, and all three Veo 3.1 IDs are preview models with more restrictive rate limits. Prices and request shapes on preview models are exactly the sort of thing that moves.

There is also an engineering dimension that tends to get discovered late. Video generation on these platforms is asynchronous — submit, get an id, poll. Vercel’s AI SDK documents a polling interval default of five seconds and a timeout default of ten minutes for gateway video jobs. BFL documents its draft mode as returning a fast preview rather than a full render, and it follows — our own inference rather than anything a vendor states — that the cheap exploration tiers are the ones least likely to strain those defaults; it is the long full-quality finishes — continuation clips, 2K regenerations, Standard-tier 4K — that push against timeouts and want the background-job flow. The cost lever and the reliability lever happen to point the same direction, which is a convenient accident rather than a designed property.

Finally, a sourcing discipline. During research for this piece we found round-up sites republishing vendor video rates with errors, including at least one figure below Google’s own published floor for the Veo 3.1 Lite tier. Read the rate off the vendor page before it goes in a budget. If you are auditing which of these models you can genuinely build on rather than merely price, our companion piece on which AI creative models actually ship as advertised is the sharper tool, and the licensing terms behind these prices deserve their own read before any client deliverable.

The operating rule is three sentences long. Generate every exploration pass on the cheap tier. Put the cheap passes in front of whoever signs off, and make the sign-off explicit — this shot, this variant. Then promote only the approved variant through the vendor’s documented mechanism, never through a fresh prompt.

The third sentence is the one teams skip, and it is where the money leaks back out. Re-prompting the winner at full quality feels equivalent and is not: on FLUX 3 the documentation is explicit that a fresh submit is its own generation and can interpret the shot differently. If you re-prompt, you have paid for the drafts and then gambled on the finish, which is the worst of both structures.

Where this discipline compounds is volume. A single hero film saves a few dollars. A catalogue programme generating hundreds of shots moves a line item, and the operational shape is the same one we described for running product video at catalogue scale — a cheap generation pass, a human selection gate, then an expensive promotion pass on winners only. If you are assembling the surrounding asset set rather than a single sequence, the model-routing decisions for wiring this into a real product-launch page sit alongside it.

Looking forward, we expect the lock-in mechanism to become a competitive axis rather than a footnote. Two of the three platforms examined here have already shipped one, and they shipped it because the alternative — charging full price for exploration — is a worse product, not because it was cheaper to build. The vendors that add a documented draft-to-final path will win the workflows that involve approval steps, which is most commercial work. The ones that only publish a cheaper model will keep winning single-shot experiments. Both are viable positions; only one of them survives contact with a client review round. Teams building the surrounding pipeline normally run this alongside a broader paid media production programme where the render bill is one line among many, and an always-on content engine is where the per-second savings actually compound.