AI video pricing index: what a second costs in July 2026

A second of AI video costs between $0.045 and $0.68, depending who you ask

The cheapest rate we found for a second of AI video generation in July 2026 is $0.045 (Hailuo 02 Standard, fal.ai, aggregator-reported). The most expensive is $0.682 (Seedance 2.0 1080p standard tier, fal.ai). Between those two numbers sits every other AI video price on the market right now, and the gap between them is a 15x spread for output that lands in a similar quality band on some prompts. That gap is the actual story of this index. Below is every rate we could find for five actively-hosted video models, each one labeled with how confident we are in it: verified against a primary source, aggregator-reported (secondary or search-synthesized), or Infer rate observed directly from the model's own catalog page. This is the July 2026 edition; we're building it to update quarterly, so the table structure below is designed to diff cleanly against the next one.

Infer hosts all of these models and profits the same whichever one you pick. This report exists to make the actual number visible, not to sell a favorite.

Key findings

• The cheapest and most expensive per-second rates for AI video generation, across every provider we checked, differ by roughly 15x: $0.045 (Hailuo 02 Standard, fal.ai) to $0.682 (Seedance 2.0 1080p standard, fal.ai).
• Resolution and audio surcharges can outweigh the base model choice: Veo 3.1 Fast on fal.ai runs $0.10/sec at 1080p without audio but $0.35/sec at 4K with audio, a 3.5x jump for one model with no vendor switch required.
• Infer's hosted rate for Seedance 2.0 Pro ($0.13/sec) is roughly a fifth of fal.ai's quoted 1080p standard tier for the same model family ($0.682/sec), though the two aren't a clean apples-to-apples match; see the resolution note below.
• Only three Infer-hosted video models have both a confirmed per-second price and a leaderboard Elo we could verify: Veo 3.1 Fast, Kling 3.0 1080p Pro, and Seedance 2.0 Pro. Everything else in the catalog is missing one half of that pair as of this pass.
• Veo 3.1 Fast delivers the best Elo-per-dollar of the three we could compute: 12,716 Elo points per dollar-per-second, ahead of Kling 3.0 1080p Pro's 12,480 and Seedance 2.0 Pro's 9,777.
• Of the 20 price points in the master table, only three met our bar for "verified against a primary, non-aggregated page" — Google's own Veo 3.1 rates, fetched directly from ai.google.dev's pricing page. That fetch also proved the point of the exercise: the aggregator consensus had quoted Veo 3.1 Fast at ~$0.15/second, and Google's actual page says $0.10–$0.12. Five rows are Infer's own displayed rate; the remaining 12 are aggregator-reported, a ratio worth sitting with before treating any competitor pricing chart as gospel, including this one.
• Three of Infer's own catalog pages (Hailuo 02 Pro, Ideogram 3.0, Imagen 4 Standard) returned HTTP 500 errors during compilation; their prices were recovered from server-rendered meta tags, not the live pricing block, and are flagged accordingly below.

Methodology

Every price below was pulled directly from a live page (Infer's own model catalog pages, fal.ai model listings, or provider documentation), or, where a direct fetch wasn't possible, from a search-aggregated secondary source, and labeled as such. We did not derive, round, or estimate any figure; where a source gave a range, we kept the range. All observations are stamped 2026-07-22, the day this data pack was compiled, four days before this report's July 28, 2026 publish date. Elo figures come from Artificial Analysis's Video Arena as cached on Infer's leaderboard page (snapshot dated 2026-04-30). Artificial Analysis's own live arena shows a materially different ranking as of this writing: Kling 3.0 1080p Pro sits at #6/1111 Elo on AA's live page versus #3/1248 on Infer's cached copy, which is exactly the kind of drift this index exists to make visible. We cite whichever snapshot we're quoting, with its date, every time.

The master price table

Every rate we found for the five video models currently hosted on Infer, across every platform we could locate a number for. "Verified" means we fetched the exact page cited and it's a primary, non-aggregated listing. "Aggregator-reported" means the figure came through a search-synthesized answer or a secondary reseller/guide site. "Infer rate observed" means it's the price displayed on Infer's own model page as of the observed date.

Two models on the leaderboard don't get a row above on purpose. HappyHorse-1.0 sits at #1 overall (1368 Elo) on Infer's cached leaderboard snapshot, but Infer's own page note says it has "no public API yet," so there's no price to quote. Runway Gen-4.5 (1216 Elo on Infer's snapshot) isn't hosted on Infer, and no per-second rate for it turned up in this pass through a source we'd cite as even aggregator-reported; we're not inventing one.

What Elo actually buys you per dollar

Price alone doesn't tell you what you're getting; pairing it with leaderboard position does. Below is Elo divided by per-second price for the three Infer-hosted models where we have both a confirmed price and a confirmed Elo from the same snapshot (Infer's leaderboard, Artificial Analysis data, dated 2026-04-30). Higher is better: more leaderboard quality per dollar spent.

Read this table carefully before quoting it. Veo 3.1 Fast technically posts the best Elo-per-dollar figure here, but its 1208 Elo is the lowest of the three in absolute terms: cheap-and-fine beats expensive-and-fine on this metric even when "fine" isn't the top result. Seedance 2.0 Pro has the highest raw Elo (1271) and the joint audio-video generation neither of the other two ships natively, so its lower Elo-per-dollar score isn't a knock on the model. It's the cost of the only model here that does audio in one pass.

Hailuo 02 Pro and Wan 2.2 T2V-A14B are both excluded from this table on purpose: neither appears with a numeric Elo entry in Infer's 12-row Artificial Analysis-sourced leaderboard (Hailuo's Infer page also 500-errored during compilation, so we don't have a confirmed spec sheet to pair with its price either). Infer's own copy calls Wan 2.2 "#1 open-weights leaderboard (Q4 2025)" and claims it delivers "~85% quality compared to Hailuo 02 Pro." Both are Infer's stated positioning, not independently-scored Elo figures, so we're not converting them into a per-dollar number we can't back up.

What four real jobs actually cost

These are computed directly from the Infer per-second rates in the master table above, with no rounding beyond the source figures and no hidden fees added.

The spread compounds with volume, not with any single job. A single 6-second clip separates the cheapest and priciest model here by $0.30, easy to ignore. Run a 100-variant test batch and the same per-second gap becomes a $30 difference between Hailuo 02 Pro and Seedance 2.0 Pro, and that's before anyone asks whether the extra $30 bought better audio sync or camera control. It's a legitimate question. Seedance's native joint audio-video generation is real functionality Hailuo doesn't claim, but it's a question worth asking with the number in front of you, not after the invoice.

Veo 3.1 Fast's 8-second native cap (chainable to roughly 148 seconds per Infer's spec sheet) means the 90-second animatic above isn't one continuous render. It's stitched from multiple generations, which is worth knowing before you budget it as a single clip.

The billing gotchas

Three things move the real cost of a render away from the sticker per-second rate, and none of them show up in the base price row.

Fixed-price tiers hiding behind a per-second quote. MiniMax's official pay-as-you-go documentation lists Hailuo 02 Pro at $0.08/sec, but the same pricing page also cites a legacy fixed-price tier: $0.28 or $0.56 for a 6- or 10-second clip at 768p, $0.49 flat for a 6-second 1080p clip, and $0.10 or $0.15 for 6- or 10-second 512p clips. Do the division on that 1080p figure. $0.49 for 6 seconds is about $0.082/sec, close to but not identical to the per-second quote, and you can see how "per-second pricing" and "fixed-tier pricing" for the same nominal product can drift apart by a few cents depending which page you're reading. Confirm which billing mode actually applies to your account before estimating a production run.

Resolution and audio multipliers, not flat surcharges. fal.ai's rate card for Veo 3.1 Fast is the clearest concrete case in this dataset: $0.10/sec at 720p/1080p without audio climbs to $0.15/sec with audio, and jumps again to $0.30/sec (no audio) or $0.35/sec (with audio) at 4K. That's a 3.5x total spread for what's nominally "the same model" depending on two toggles. If a vendor quotes you a headline per-second number, ask which resolution and audio setting it assumes. The answer is rarely the cheapest one.

Failed-generation policy: unresolved in this pass. Whether a provider bills you for a render that errors out or fails a content filter is a real cost variable at scale, and it's the one gotcha we can't give you a sourced number for. We could not confirm Infer's specific failed-generation billing policy from a primary source during this compilation. Check the console documentation for your specific account before assuming either outcome. It's exactly the kind of detail that's invisible until a batch job fails at scale.

Predictions

Video pricing compresses further before it diversifies again. The 3.5x resolution/audio multiplier fal.ai already applies to Veo 3.1 Fast is a preview of where the rest of the market goes: base per-second rates keep sliding toward Hailuo 02 Standard's $0.045/sec floor, while providers claw margin back through resolution, audio, and reference-image tiers instead of the headline number. Expect the "verified" column in this index to get harder to fill, not easier. As more providers gate real pricing behind consoles and account tiers (the way tryinfer.com/pricing currently 404s, per our July 2026 compilation), the gap between a public headline rate and what an account actually pays will widen. We'll revisit both predictions in the next edition of this index.

Chinese-developed models (Kling, Seedance, Wan, Hailuo) will keep undercutting the Google/OpenAI-adjacent rate card on a pure per-second basis. The audio-native tier, Seedance's joint generation and Veo's Foley-quality output, is where price convergence is least likely; that functionality currently commands a real premium, not a rounding difference, and nothing in this dataset suggests that's about to change.

Cite this report

Infer Team. "AI video pricing index: what a second costs in July 2026." Infer, July 28, 2026. https://tryinfer.com/guides/ai-video-pricing-index

Run these numbers yourself before you commit a budget: Veo 3.1 Fast, Kling 3.0 1080p Pro, Seedance 2.0 Pro, and Wan 2.2 T2V-A14B are all live on Infer under one API key. Hailuo 02 Pro is also hosted on Infer, though its catalog page was 500-erroring at the time of this compilation; try it directly to confirm current status.

For the models behind these numbers, see Kling 3.0 vs Veo 3.1: which wins in 2026?, Seedance 2.0 vs Kling 3.0: the top of the leaderboard, tested, Wan vs Kling vs Hailuo: China's video models, compared, the best AI video generation models in 2026, ranked, cheapest AI video API, and state of AI video 2026. Full catalog and leaderboard context lives on the guides hub.