AI Video Price War: What 2K at $0.13/sec Means for Ads

The AI video price war reached ad-creative economics on July 31, 2026, when MiniMax’s H3 video model and ByteDance’s Seedance 2.5 launched on the same calendar day. H3 shipped with published pay-as-you-go pricing — $0.13 per second at 2K resolution — which makes a 30-second ad variant roughly $3.90 to generate, computed directly from the vendor’s stated rate.

That number lands differently for marketers than for developers. Most coverage of these launches repeats per-second rates in the abstract; almost none does the arithmetic to a real ad deliverable — a 6-second bumper, a 15-second social cut, a 30-second in-stream spot. When the math is done honestly, generation cost stops being the constraint on creative-testing volume. Human review throughput, brand-safety sign-off, and copyright posture become the constraints instead.

This cost playbook covers what actually launched, which prices are vendor-published versus third-party estimates, a recomputed cost table across finished ad-asset durations, the copyright backdrop that should shape routing decisions, and a practical framework for deciding which model gets which workload. Every dollar figure below is labeled published or estimated, and every derived cell is recomputed from the stated per-second rate.

MiniMax released H3 on July 31, 2026, after a WAIC tease on July 17 — a video-generation model, successor to the Hailuo line, and distinct from the MiniMax M3 text model that shipped in June. MiniMax’s own docs brand it only “MiniMax H3” (model ID MiniMax-H3); “Hailuo 3.0” is a community label, not an official name. Reuters coverage lists advertising, e-commerce, product design, and games as the target commercial use cases. For the full capability teardown — 2K output at 24fps, native synced audio, reference-to-video — see our dedicated H3 launch coverage.

The same day, ByteDance’s Seed team published its own launch post, opening plainly: “Today, we are officially launching Seedance 2.5.” The model went live on Jimeng Web and Doubao Pro — China consumer apps first — with the BytePlus ModelArk developer API listed as coming soon. We break down the one-take generation and 50-input referencing system in our Seedance 2.5 launch analysis.

The same-day collision matters more than either release alone. Each vendor’s individual coverage will dominate search results for its own model name; what neither frames is that China’s video price war — running since Seedance 2.0 and the Hailuo line began undercutting Western per-second rates — now lands directly on the line item marketers actually budget: cost per finished ad asset. That is the lens for everything below.

Exactly two vendors in this comparison have per-second rates published on their own pricing pages as of July 31, 2026. MiniMax lists H3 at $0.13 per second at 2K and $0.09 per second at 768P — with two footnotes that matter for budgeting. First, H3 is pay-as-you-go only; prepaid Hailuo token packages explicitly do not support it yet. Second, the 768P tier is closed beta — the pricing page says access requires contacting sales, so the $0.09 rate is not self-serve today. Input materials add small increments: audio input is free, images are free for the first 5 per project then $0.04 per additional image, and video-input material bills at the same per-second generation rates.

Google’s Gemini API pricing page, fetched the same day, lists Veo 3.1 in three tiers: Standard at $0.40/second for both 720p and 1080p (video with audio is the default, included in that rate) and $0.60/second at 4K; Fast at $0.10/second for 720p; and Lite at $0.05/second for 720p, with no 4K on Lite. The prior-generation Veo 3 was deprecated with a June 30, 2026 shutdown, so Veo 3.1 is the only current production tier.

Per-second rates are abstractions; media teams budget in deliverables. The table below recomputes every cell from the stated per-second rate multiplied by the duration — no cell is copied from a third party’s pre-computed total, and only vendor-published rates get a column. Durations map to real ad formats: 6-second bumpers, 15-second social cuts, 30-second in-stream spots, and 60-second hero or landing-page loops.

*H3’s maximum single clip is 15 seconds, so the 30-second and 60-second H3 cells assume multiple stitched clips — billing is per-second, so the cost scales linearly, but the workflow includes an assembly step that Veo’s longer single renders do not require. Read the comparisons as cost-only, not quality-equivalent: 2K and 1080p are different resolutions, and none of these outputs has been benchmarked head-to-head for ad-creative fidelity.

The spread is still striking. A 30-second variant computes to $3.90 on H3’s 2K rate versus $12.00 on Veo 3.1 Standard at 1080p — roughly 3× H3’s cost for the same duration — and $18.00 at 4K. Run a 20-variant hook-testing matrix of 15-second cuts and the gap compounds: $39.00 on H3 2K versus $120.00 on Veo Standard 1080p. At these absolute levels, either bill is a rounding error next to the media spend the variants feed — which is precisely the point of Section 05.

Seedance 2.5 launched with no price at all — ByteDance’s post carries no pricing section, a detail we confirmed on a direct fetch of the launch page. The only Seedance dollar figures in circulation are third-party reseller estimates from the Seedance 2.0 era — roughly $0.10 to $0.80 per minute, per a February 2026 aggregator comparison — and they predate 2.5 entirely. They are directional context for “China video pricing tends to undercut Western labs,” nothing more. For how the model itself stacks up creatively, see our Seedance 2.5 vs FLUX 3 and Gemini Omni field comparison.

Kling 3.0 and Runway Gen-4.5 sit in a similar evidentiary gap. Aggregator sources converge on roughly $0.084–$0.168 per second for Kling 3.0 depending on tier, and roughly $0.12–$0.15 per second for Runway Gen-4.5 — but neither range could be verified on the vendor’s own pricing page during this research pass (both pages are JS-gated to scrapers), so we present them as reported, not published, and exclude them from the cost table above. Directionally, both land in the same band as Veo’s Fast tier and H3’s published rates — the per-second price war is a field-wide phenomenon, not a two-vendor stunt.

Quality signals are thinner still. No independent benchmark or arena score exists for H3 or Seedance 2.5 as of July 31 — H3’s promised open weights were “within days” away at launch, a promise rather than a shipped fact. The predecessor Seedance 2.0 topped Artificial Analysis’s Video Arena blind human-preference leaderboard in pre-2.5 rankings, per aggregated arena data — a result that belongs to 2.0 and must not be projected onto 2.5. Note also that Artificial Analysis’s video pricing figures are a standardized cost simulation (USD per minute for a 720p, 5-second video, built from each provider’s published rate), not an independent measurement of real workloads.

Here is the trend worth interpreting rather than just reporting. When a 30-second 2K variant costs $3.90 to generate, the generation line item is cheaper than the human time spent reviewing it. The scarce resource in a creative-testing program stops being render budget and becomes everything wrapped around the render: human QA throughput, brand-safety sign-off, legal review of what the model actually produced, and the trafficking work of getting variants live. Teams that treat cheap generation as a license to 10× their variant count without scaling review capacity will simply move the queue from the GPU to the approvals inbox.

The practical response is to industrialize the review layer the same way generation just got industrialized: tiered QA (automated checks for logo fidelity, claim compliance, and aspect-ratio safe zones before any human looks), sampling-based review for low-stakes variants, and full human sign-off reserved for hero assets. This is the same storyboard-first discipline we laid out in our guide to building an ad-creative pipeline around Seedance, and it pairs naturally with product-video localization at near-free per-second rates for ecommerce catalogs.

For paid-media teams specifically, the arithmetic changes what a healthy testing cadence looks like. A $39.00 twenty-variant matrix against a five-figure monthly media budget means creative testing volume is now bounded by how fast you can read results and kill losers — a measurement and workflow problem, which is exactly where a disciplined paid-media operation earns its keep.

The per-second math is only half the routing decision. On February 22, 2026, the Motion Picture Association sent its first-ever AI cease-and-desist letter over video generation, targeting ByteDance’s Seedance 2.0 for what it called systemic infringement of copyrighted Hollywood content — days after viral AI-generated clips of Brad Pitt and Tom Cruise circulated. Disney, Paramount, Netflix, Warner Bros., and Sony each separately sent their own letters over unauthorized use of Marvel, Star Wars, Star Trek, South Park, and Stranger Things characters.

ByteDance responded by voluntarily pausing Seedance 2.0’s global rollout in March 2026 and pledging safeguards. That history is why the July 31 launch sequence looks the way it does — enterprise beta in June, a narrower API opening on July 16, then a China-apps-first consumer launch — and it is the concrete reason enterprise buyers weigh legal exposure alongside per-second cost. A brand-hero spot that ships to broadcast carries indemnification questions that a $3.90 generation bill does not answer. Google’s enterprise licensing posture and published terms are a real part of what the 3× Veo premium buys, wholly apart from output quality.

Projecting forward: expect the price war to keep compressing per-second rates while the legal layer becomes the durable differentiator. If MiniMax ships H3’s promised open weights, and if Seedance 2.5’s API arrives with published pricing and cleaner provenance controls, the routing calculus below will need re-running — but the structure of the decision (stakes up, legal posture up) should outlast any individual rate card.

The workable framework assigns each ad workload to a model class by what failure costs, not by what generation costs. It is the same logic as our broader framework for routing AI spend by task stakes, applied to video.

Two disciplines make the framework hold. First, re-verify rates at budget time — the 768P beta gate, the pay-as-you-go-only restriction on H3, and Veo’s tier structure are all details that change unit economics if missed. Second, keep the published-versus-estimated distinction alive inside your own planning docs the way this post does: the moment a third-party estimate hardens into a budget line, you have built a forecast on a number nobody will stand behind.