AI Video Model Pricing, August 2026: Seedance, Kling, Veo, and Wan, Which Is Cheapest

If you remember one number from this piece, make it "0.022 to 0.60" — the spread, in dollars per second, between the cheapest tier we found (Seedance 2.0 Fast, via third-party sources) and the most expensive (Veo 3.1 Standard at 4K, straight from Google's own rate card): nearly 30x. And that's before accounting for the extra 20%-plus you can lose or save just by picking the wrong relay platform for the same exact model. Video generation pricing is a genuinely messier problem than text-model pricing — per-second, per-token, per-credit, and per-clip billing all show up across these four vendors, and "what does it actually cost per second" turns out to be a question you have to work to answer rather than read off a single number.

This piece covers four of the most-watched video generation models right now: ByteDance's Seedance 2.0 and 2.5, Kuaishou's Kling 3.0, Google's Veo 3.1, and Alibaba's Tongyi Wanxiang (Wan) 2.7. Our method has two layers: first, pull whatever raw rates each vendor's own documentation or pricing page actually publishes; second, cross-check real quoted prices from multiple third-party relay and aggregator platforms — including several eggstriker has already reviewed in depth, like Atlas Cloud, Krea, Getimg.ai, WaveSpeedAI, and Kling's own official developer platform. Wherever a number isn't published anywhere official and only exists as a third-party estimate, we say so plainly rather than smoothing the table into something that looks tidier than the underlying reality.

1. The Full Picture: All Four Models' Pricing, Side by Side

Here are the core numbers. In the "Official Pricing" column, we've quoted the vendor's own rate card wherever we could find one; where we couldn't, we say so instead of forcing a number that isn't really there. The "Relay / Third-Party Reference Price" column reflects figures that showed up consistently across multiple independent sources — useful for understanding the general price range, but not a substitute for checking a live quote before you commit.

One thing worth flagging outside the table: Seedance 2.5's official pricing dropped on the very day we wrote this, so relay platforms almost certainly haven't had time to onboard it or set their own rates yet. We checked Replicate's model page directly and confirmed it currently reads "coming soon" — if you want to reach 2.5 through a relay right now, you're probably still waiting. More on that in Section 9.

2. ByteDance Seedance — 2.0's Two Tiers, and 2.5's Pricing Landed Today

Seedance 2.0 has been live for nearly half a year now (launched February 12, 2026), sold through Volcano Engine Ark and billed along four axes: resolution (480p/720p/1080p/4K), duration (4–15 seconds), generation mode (Standard/Fast/Mini), and whether audio or video reference inputs are included. Third-party aggregations put the cheapest Fast tier at roughly $0.022/sec and the priciest 4K tier at around $0.353/sec — a 16x spread inside the same model, which is a good reminder that "how much does Seedance cost per second" isn't a single-number question until you specify the tier and resolution.

The real news is Seedance 2.5: officially launched July 31, 2026 (eggstriker already has a dedicated review breaking down its feature upgrades — single shots doubling from 15 to 30 seconds, reference inputs expanding from roughly 15 to 50), but at the time, developer-facing API pricing was a total blank, with only a "coming soon via BytePlus ModelArk" line to go on. That gap closed on August 7, 2026 — the day we're publishing this — when Volcano Engine posted token-based pricing rather than a straight per-second rate: ¥70 per million tokens without video input, ¥42 per million with video input, with official examples of roughly ¥3.36 for a 5-second 480p clip and ¥7.56 for 5-second 720p. Converted to a rough per-second figure, that's about $0.10/sec at 480p and $0.23/sec at 720p — roughly 40-50% above Seedance 2.0's official Standard tier at the same resolution, which lines up reasonably well with a video model that just doubled its shot length and more than tripled its reference-input ceiling.

On quality, Seedance 2.0 briefly led Artificial Analysis's text-to-video Elo leaderboard at launch and currently sits at #3 with a score of 1,225, passed by MiniMax H3 (1,242) which launched the same day as 2.5. It's still considered top-tier on multi-shot consistency, camera control, and physics simulation. Seedance 2.5 doesn't have an independent third-party benchmark yet as of publication — the feature upgrades in ByteDance's own blog post are real product specs, but exactly how much better 2.5 performs is still self-reported for now, a point we've already covered in more depth in our dedicated Seedance 2.5 review.

3. Kuaishou Kling 3.0 — Official Pricing Runs on Credits, Not Dollars

Kling has one of the fastest release cadences in the video-generation space, running from v1 through v2.5 and now an O1 reasoning-capable version. One detail worth flagging: eggstriker's own official review of the Kling AI developer platform (klingai.com), last verified August 6, 2026, lists the product lineup as v1 through v2.5 plus O1 — it does not mention a "3.0" version anywhere. Meanwhile, multiple third-party pricing trackers we checked this round, plus our own site's Krea and Atlas Cloud reviews, consistently use "Kling 3.0" or "Kling V3.0," and an official-looking API guide document is dated February 6, 2026. We couldn't find a clean official explanation reconciling the two naming conventions. This piece presents pricing under the widely-used "Kling 3.0" label, but if you go check the official site directly, the version naming you see there may not line up exactly — confirm the actual model ID on klingai.com/dev before you build against it.

Official pricing runs on a credit system rather than a straight dollar-per-second rate: 6 credits/sec for no-audio 720p, 8 credits/sec for no-audio 1080p, 9 credits/sec for native-audio 720p, 12 credits/sec for native-audio 1080p, plus a 2-credit/sec voice-control add-on. Against the entry-level $5-for-330-credits pack, that works out to roughly $0.09–$0.18/sec depending on resolution and audio. This lines up well with a separate third-party summary putting official pricing at "$0.084–$0.168/sec (Standard to Pro mode)" — the two figures cross-check cleanly.

Worth noting: Kling's official pricing page is genuinely harder to find than most — the full rate table sits behind a developer console login, and most of what's publicly searchable comes from third-party aggregation rather than a scraped official page, which is why we didn't give Kling the "most transparent" label (that goes to Veo 3.1 in the next section). On the relay side, Krea quotes Kling 3.0 as low as $0.05/sec, undercutting even the official entry tier, and other resellers like EvoLink quote around $0.075/sec. This "official costs more than third-party" pattern echoes something we've seen before with open-weight text models like GLM-5.2, but the mechanism is different here: Kling is a closed model, so a cheaper third-party price more likely comes from bulk purchasing discounts or how credit bundles get split up internally, not genuine compute-level price competition the way it works with open weights.

On the quality side, Kling 3.0 holds four of the top ten spots on Artificial Analysis's text-to-video leaderboard (across different resolutions/modes) — the most leaderboard visibility of the four models we cover here. Its officially claimed differentiators are hitting 4K/60fps in a single 15-second shot, and fine-grained lip-sync across multiple characters in the same scene — both points that have held up well in Chinese creator-community feedback.

4. Google Veo 3.1 — The Most Transparent Rate Card, and the Most Expensive

Compared to the previous two, Veo 3.1 is the only model in this comparison where you can open the official docs and read a complete pricing table without decoding a credit system or a token conversion: Google's Gemini API pricing page lists three tiers and six concrete figures directly — Lite at $0.05/sec (720p) and $0.08/sec (1080p, no 4K support); Fast at $0.10/sec (720p), $0.12/sec (1080p), and $0.30/sec (4K); Standard at a flat $0.40/sec for both 720p and 1080p, and $0.60/sec at 4K. All of these already include audio generation, and Google explicitly notes you're only billed if the video generates successfully — a failed audio-processing step, for instance, doesn't get charged.

The tier structure itself sends a clear signal: Google split "do you need top-tier quality" and "do you need faster turnaround" into two separate axes. Lite targets budget-sensitive use cases, Fast is the sweet spot for most production work, and Standard is aimed at professional workflows where quality matters more than budget. Lined up against the other three, Veo 3.1 Standard at 4K ($0.60/sec) is the single highest official rate we found across this whole comparison — roughly 70% above Seedance 2.0's Standard tier, and two to three times Kling 3.0's official entry rate. But if you only look at Lite ($0.05/sec at 720p), the price actually sits in the lower range among all four — the "Veo is expensive" impression is really about the Standard/4K tier specifically, and picking the right tier makes a real difference.

On quality, multiple third-party reviews consistently describe Veo 3.1 as the most well-rounded all-rounder — native audio, dialogue lip-sync, and prompt adherence are widely regarded as the strongest in its class, and its developer usability and documentation are the most mature of the four, something we've also cross-verified in our earlier Seedance 2.5 review. For scenarios that need accurate dialogue audio-video sync on the first pass — think character-driven short-form drama or ad script previews — Veo 3.1 is currently still the safest first choice, even at the premium.

5. Alibaba Wan 2.7 — The Only Open-Weight, Self-Hostable Option

Alibaba's Tongyi Wanxiang (Wan) 2.7 launched April 3, 2026, and it's positioned differently from the other three: the public DashScope / Model Studio docs only note that "resolution affects cost" and "billing is per second — check model pricing before calling," without listing an actual rate card the way Veo does. You need to log into the console per model to see a number — similar friction to Kling's official pricing.

But Wan 2.7 has one property none of the other three share: its image and video weights are released under Apache 2.0, distributed primarily through Alibaba's own ModelScope with GitHub as a secondary channel (a few third-party sites note that the official Hugging Face organization still only hosts the previous 2.1/2.2 weights, so if you're planning to pull weights from Hugging Face specifically, double-check the version number first). That means teams with GPU access can, in principle, bypass every API and relay layer entirely and self-host — third-party assessments put the practical floor at around 16GB VRAM, ideally 24GB, a noticeably lower bar than most self-hosted text models.

Because the weights are open, third-party hosting competition is more genuine for Wan than for the other three: OpenRouter quotes $0.10/sec, Atlas Cloud starts from $0.10/sec as well, and EvoLink lists $0.086/sec at 720p and $0.144/sec at 1080p; the previous-generation Wan-2.2 Turbo Spicy runs as low as $0.026/sec on Atlas Cloud — the lowest per-second price we found anywhere across the model families in this comparison, just above Seedance 2.0 Fast's $0.022/sec and well below the entry-level official rate of any of the other three. This mirrors a pattern we've repeatedly verified with open-weight text models (DeepSeek, GLM, Kimi): open weights let third-party inference providers compete on price, and the official rate card is frequently not the market's lowest.

On capability, Wan 2.7's most distinctive feature is "Thinking Mode" — before generating the actual frames, the model plans out composition, camera blocking, and spatial layout first, aimed at improving coherence in complex scenes with multiple characters or complex camera movement. It's the only one of the four to explicitly call out a separate planning step before generation. It also supports clips up to 15 seconds, up to 5 reference images for multi-character consistency, and start/end-frame control — a feature set aimed at a similar goal as Seedance 2.5's "multi-round extension plus 50 reference inputs," but Wan makes the generation process itself an explainable intermediate step rather than simply feeding in more reference material.

6. Choosing on Quality: Elo Scores, 4K, Duration, and Lip-Sync

Lining up all four vendors' differentiators side by side, it's clear each one is betting on a different sub-dimension rather than competing head-on in the same lane:

• Seedance 2.0/2.5: Multi-shot consistency, physics simulation, native 4K plus unified audio-video generation, with 2.5 doubling down on 30-second single shots and multi-round extension — best for content that needs long-form narrative coherence.
• Kling 3.0: Highest leaderboard visibility on Artificial Analysis (4 of the top 10 spots), single 15-second shots can hit 4K/60fps directly, and standout multi-character lip-sync — the strongest track record in Chinese creator communities.
• Veo 3.1: Widely regarded as the strongest for dialogue audio-video sync, prompt adherence, and developer usability — the first choice when accuracy on the first pass matters, at the cost of the highest official pricing.
• Wan 2.7: The explainability of Thinking Mode plus the pricing flexibility that comes with open weights make it the model with the most savings paths of the four, though it currently has less independent-benchmark visibility than the other three.

One thing worth stating honestly: specific Elo numbers vary noticeably across different third-party tracking sites (for instance, some sites' Kling scores don't line up with the Artificial Analysis figures we cross-checked directly). This piece only cites claims that multiple independent sources corroborate (like "Kling 3.0 holds four of the top ten leaderboard spots" or "Seedance 2.0 sits at #3 with 1,225 Elo"). Where a specific single-source number couldn't be cross-verified, we chose not to cite it rather than pick whichever number looked most flattering.

7. Two Access Routes: Direct Vendor APIs vs. Aggregator Relays

For developers, there are basically two paths into these four models.

Route one: direct vendor APIs. ByteDance runs through Volcano Engine Ark / BytePlus ModelArk, Kuaishou through the Kling AI developer platform (klingai.com/dev), Google through the Gemini API / Vertex AI, and Alibaba through DashScope / Model Studio. This route typically has the lowest raw pricing (no relay markup) and is the only route that guarantees first access to the newest models the day they launch — as with Seedance 2.5's pricing today. The tradeoff is integrating four completely different API protocols, auth schemes, and billing logics separately, with billing scattered across four different consoles.

Route two: an aggregator/relay that unifies multiple models under one account. fal.ai, Replicate, OpenRouter, and vendors eggstriker has already reviewed — Atlas Cloud, Krea, Getimg.ai, WaveSpeedAI, and others — all offer a single account and API protocol covering two or three of these four models simultaneously (very few platforms cover all four; see the next section for specifics). The value here is skipping the repeated integration work and being able to compare model outputs side-by-side in the same playground, but it typically costs somewhat more than the direct vendor rate (standard closed-model resale economics), and new models like Seedance 2.5 always show up with some lag — days to weeks — before a relay lists them.

8. Same Model, Different Platform: What Seedance 2.0 at 720p Actually Costs

Route two has real price competition baked into it. Take Seedance 2.0 at 720p with video input as an example — quotes we cross-checked came out to Atlas Cloud at $0.1486/sec, WaveSpeed at $0.15/sec, OpenRouter at $0.1512/sec, and fal.ai at $0.1814/sec — a roughly 22% spread between the cheapest and most expensive. These figures largely trace back to Atlas Cloud's own comparison blog post, and as a vendor comparing itself favorably against competitors, that kind of self-published comparison content naturally carries some promotional bias. We're presenting it here as a rough sense of price magnitude, not an endorsement of any one platform — check each platform's live, current quote before committing.

The lesson here is simple: even for the exact same model and spec, a roughly 20% price spread between relay platforms is real. For high-volume production (say, a content team burning through hundreds of thousands of seconds of video a month), that gap adds up to real money, and it's worth shopping around a few platforms before committing to production integration.

9. Which of eggstriker's Reviewed Vendors Actually Sell These Four Models

This is the question eggstriker readers care about most: of the dozen-plus multimodal vendors already reviewed on this site, who actually sells Seedance, Kling, Veo, and Wan? We checked each vendor's model catalog directly, and here's what we found:

A few things worth calling out. First, Getimg.ai is the only vendor we checked whose catalog covers all four target models at once, but its API access and web subscription are two separate systems — before you integrate, confirm you're getting an API key that can actually call these four video models, not just a web-only subscription credit. Second, Atlas Cloud's video lineup explicitly includes Seedance 2.0, Kling V3.0, and Wan 2.7, but Veo doesn't show up anywhere in its catalog — if Veo 3.1 is a hard requirement for you, Atlas Cloud isn't currently a complete one-stop option. Third, the Kling AI developer platform is the only genuinely "direct official" option on this list — RMB billing, no proxy needed — but as its name suggests, it only sells its own model, not resold access to Seedance, Veo, or Wan. Fourth, as of publication, Seedance 2.5 hasn't been confirmed live on any relay platform we checked — Replicate's model page directly reads "coming soon" — which is the flip side of the same phenomenon we flagged in Section 2: pricing just landed, and relays haven't caught up yet.

10. What Matters Most for Mainland China Readers: Who You Can Reach Directly

There's a distribution pattern across these four models that's easy to overlook: Seedance (ByteDance), Kling (Kuaishou), and Wan (Alibaba) are all products of mainland Chinese companies, and their official channels — Volcano Engine, the Kling AI developer platform, DashScope — were built natively for mainland Chinese users, billed in RMB, no proxy required. Veo is the exception: it's a Google product, and no matter which access route you choose — Google's own Gemini API/Vertex AI, or any overseas aggregator — mainland China users can't reach it without a proxy or dedicated line. It's the one model in this comparison where switching vendors doesn't remove the access barrier.

Practically, that means if your product or workflow is sensitive to both budget and access friction, and you can live without Veo, you can in principle reach Seedance, Kling, and Wan entirely through their respective official channels without touching a relay at all, at first-party official pricing (Wan 2.7's DashScope console rate, for instance, should in theory undercut any third-party relay reselling it, since a reseller always needs to preserve its own margin). On the flip side, if you genuinely need Veo 3.1 — say, for its dialogue audio-video sync — the proxy/dedicated-line layer of cost and risk doesn't go away no matter which route you pick, and at that point a relay's value shifts from "convenience" to "peace of mind": for most teams, paying an aggregator that's already solved the Google-facing infrastructure problem is more cost-effective than building and maintaining that proxy chain yourself.

There's one more angle worth flagging for mainland teams: Wan 2.7's open weights carry extra value here. If you have GPU access meeting the 16–24GB VRAM floor, you can in principle skip every API and relay layer entirely and pull the weights straight from ModelScope to self-host — a route that never touches proxying or cross-border billing at all, making it the only truly "zero-relay" option among the four, at the cost of owning your own deployment, ops, and hardware.

11. Picking by Use Case: Five Reader Profiles

12. A Note on How Fresh This Data Is

13. Conclusion

Having gone through all four vendors' pricing, an honest conclusion is this: video generation models haven't converged on a unified pricing language the way text models have with "per million tokens." Per-second, per-token, per-credit, and per-clip billing all coexist, and even a single model splits into multiple price tiers depending on resolution, duration, and whether audio is included. That means "which model is cheapest" doesn't have a single universal answer — you have to pin down the specific use case (resolution, duration, audio or not) before you can genuinely compare.

If you need one general direction: if budget is the top priority, Seedance 2.0 Fast and the previous-generation Wan-2.2 Turbo represent the current price floor; if quality is the top priority, Veo 3.1 and Kling 3.0 each lead in different ways — audio-video sync and leaderboard visibility, respectively — but both come at a real cost premium; and if what you value most is long-term independence and not being locked into any one relay platform's or vendor's future price hikes, Wan 2.7 is currently the only answer that supports true self-hosting. And for mainland China readers specifically, the simplest rule of thumb worth remembering: Seedance, Kling, and Wan can all be reached directly through official mainland channels, billed in RMB — Veo is the only one of the four that genuinely can't be reached without a proxy, a distribution pattern that may be worth remembering longer than any single per-second price point in this whole comparison.