The AI video sector is sparking a price war. Bytedance’s Ji Meng AI has tripled its pricing for end-users, with the cost of a 15-second video surging from $0.65 to as high as $5.03, while simultaneously launching a discounted version, Seedance Mini, for B-end clients; Keling AI has secured nearly $3 billion in funding and introduced promotional discounts; Alibaba’s HappyHorse entered the market at $1.20 per second and has since lowered its price to $0.72 per second. As compute costs hit rock bottom, the market will split into two extremes—those prioritizing performance premiums and those sensitive to cost—with competition shifting toward ecosystem integration and depth of workflow embedding.

Article author and source: DoNews

One side is "expensive," the other is "at rock-bottom prices."

Since July, the AI video sector has seen a series of major developments: Kuaishou’s Keling AI completed its first independent funding round of nearly $3 billion, pushing its post-money valuation to $18 billion and setting a new global record for the largest single funding round by an AI video large model company, while also establishing a firm deadline to complete its IPO before 2031. Almost simultaneously, Aishi Technology secured a C-round funding of RMB 2.98 billion, with tech giants such as Alibaba, Tencent, and Baidu jointly investing—an unprecedented alignment that has propelled the already fiercely competitive AI video battlefield into a new phase characterized by high valuations, heavy investments, and intense competition.

Amid the capital frenzy, the AI video sector is witnessing an extreme pricing divide: on one side, ByteDance’s Ji Meng AI has raised its prices three times in a row, earning user jokes about it being “too expensive”; on the other, players like Keling and Wanxing Technology are continuously cutting prices to compete for market share with rock-bottom rates.

As the technological window in the AI video industry continues to shrink, computing costs remain high, and user demands become increasingly segmented, major players are being forced to diverge—either leveraging technological advantages to command premium pricing and lock in high-value users, or relying on low-price strategies to capture market share and ecosystem positioning. When price wars hit the hard limit of computing costs, where will this competition ultimately end?

01. Byte's "Price Split": C-end prices rise, B-end prices halved

The day after Sora shut down, Ji Meng AI, under ByteDance, quietly initiated a price hike. The first round of adjustments eliminated the 40% discount for premium members, increasing the points required to generate a 15-second video from 45 to 120 points—raising user costs by 2.6 times. A flood of complaints from creators poured into the support system, prompting the platform to urgently reinstate the 40% discount within 12 hours. However, the baseline point consumption had already been permanently raised.

This is just the beginning of the series of price adjustments.

On April 3, Dream launched a dedicated VIP priority access channel for Seedance 2.0, introducing a secondary payment on top of the existing membership system: generating a 15-second video consumes 210 points, specifically designed for heavy creators unwilling to wait in line. On April 8, the platform rolled out a comprehensive price adjustment plan, reducing both membership benefits and point allocations: the first-year discount for the three membership tiers—659 yuan, 1,899 yuan, and 5,199 yuan—was tightened from 50% to 60%, and monthly point allocations were significantly cut—premium members’ monthly allocation dropped from 15,000 to 6,160 points, standard members’ from 4,000 to 2,210 points, and basic members’ from 1,080 to 725 points.

AI creators with extensive experience in short-form video have fully calculated the costs: before discount adjustments, the total cost for a 15-second video is approximately RMB 0.65; after multiple price adjustments and the addition of priority access channels, the cost can rise to as high as RMB 5.03. For a two-minute short video, the cost surges from RMB 5.2 to RMB 40.2, representing an overall increase of nearly eightfold.

The core reason for the price increase is that the computing power simply can't keep up. On the first day of Seedance 2.0's full launch, queue times quickly surged to 8–10 hours, with some users waiting as long as 24 hours. Video generation is a "bottomless money pit"—producing just a few seconds of video consumes GPU computing power equivalent to what ChatGPT would need to answer thousands of complex questions. With limited computing supply and unlimited demand, a price increase was almost inevitable.

While Ji Meng is "harvesting" on the C-side, ByteDance's other hand is significantly lowering prices on the B-side.

On June 15, Volcano Engine, a subsidiary of ByteDance, officially launched the Seedance 2.0 Mini video generation model. According to official data, the cost for image-to-video generation is RMB 0.023 per thousand tokens, and video-to-video generation is RMB 0.014 per thousand tokens. For example, generating a 720p resolution video costs approximately RMB 0.5 per second—half the price of the standard Seedance 2.0 version.

Looking at these two things together, ByteDance's logic is not contradictory.

Ji Meng targets C-end subscribers and creators with a “curate and monetize” strategy, raising prices to filter out price-sensitive users and retain a core group with strong willingness to pay. In contrast, Seedance Mini targets B-end API users and developers with a “capture and scale” strategy, undercutting industry pricing to just 0.5 RMB per second to drive higher usage volumes and deepen ecosystem integration.

However, the price reductions for B-end customers may also be due to external factors.

02. The Siege of the Price-Cutting Camp

Just as Jianmi raised its prices, several other companies were busy moving in the opposite direction.

On April 1, Keling AI announced a three-month membership discount program: Platinum-level and above individual and team members receive a 20% discount on Inspiration Points, while Gold-level members receive a 10% discount. The promotion is available on both the web and mobile app in China and runs until June 30. Meanwhile, the platform has extended the discount period for the image model, with certain features available at no cost.

Ke Ling explicitly stated in the announcement that this promotion aims to "empower every creator to create freely with stable, reliable service and competitively transparent pricing"—a polite phrasing with a clear underlying message: Ji Meng raised prices, and I’m stepping in to pick up the slack.

The head of Wanzing Studio said: "What has stalled most teams in AI animated dramas is not technology, but whether costs can support frequent updates." This precisely highlights the core dilemma in commercializing video large models—technology can be impressive, but creative teams are focused on the bottom line.

What truly intensified the competitive pressure ByteDance felt was a surprise move from Alibaba.

On April 8, the same day Yi Meng announced a comprehensive price adjustment, an anonymous developer’s model, HappyHorse-1.0, topped the Artificial Analysis video generation leaderboard, surpassing Seedance 2.0 in both text-to-video and image-to-video (no audio) categories.

Artificial Analysis employs an Elo blind test mechanism, where users vote solely based on visual perception without knowing the model identities. As of noon on April 9, HappyHorse-1.0 achieved an Elo score of 1378 in the text-to-video (no audio) track, leading second-place Seedance 2.0 by a full 105 points; in the image-to-video (no audio) track, it achieved an Elo score of 1411, setting a new all-time high on the platform.

On April 10, the Alibaba ATH (Alibaba Token Hub) business group was officially established. HappyHorse originates from the AI Innovation Division under ATH, led by team leader Zheng Bo, whose core members include Zhang Di, former VP of Kuaishou and technical lead of Keling. ATH itself is a new organizational structure launched on March 16, 2026, personally spearheaded by CEO Wu Yongming, integrating the Tongyi Lab, MaaS business line, Qwen Division, Wukong Division, and AI Innovation Division. Its guiding principle is succinctly expressed in nine characters: “Create Token, Deliver Token, Apply Token.”

If the price increase was an active choice by Jimeng based on its own technological advantages, then the emergence of HappyHorse reveals a harsher industry reality: in the video large model space, the window for technological advantage is rapidly shrinking. When a competitor can surpass you on a widely recognized blind test leaderboard using anonymity, how deep is your so-called “technological moat” really?

Alibaba quickly played its second card—lowering prices.

On April 27, HappyHorse 1.0 launched its gray test, gradually opening to enterprise creators and general users in tiers. On June 22, Alibaba released HappyHorse 1.1, systematically upgrading five key dimensions: dynamic expressiveness, subject consistency, instruction following, visual quality, and audio capabilities. Pricing is set at RMB 0.9 per second for 720p (RMB 0.54 after discount) and RMB 1.2 per second for 1080p (RMB 0.72 after discount), representing a 25% reduction compared to the 1.0 version.

In May, Wondershare FilmFactory, the AI comic creation platform under Wondershare, announced a limited-time 50% discount on its deeply integrated HappyHorse 1.0 large model, valid from May 26 to June 20. After the adjustment, the price for generating 720P video dropped to approximately RMB 0.71 per second, and 1080P video to about RMB 1.19 per second, nearing the lower end of the current AI video generation pricing range. During the same period, FilmFactory also offered varying discounts on models including Seedance 2.0, Vidu Q3, and Wan 2.6.

Ji Meng's price increase is predicated on "scarce computing power" and "technological leadership." However, when Alibaba demonstrated that it possesses superior algorithms (topping the rankings) and is willing to open its API at a significantly lower price (halved pricing), ByteDance's reliance on the "technology premium" logic has been challenged.

03. The End of the Price War

Some rise, some fall—amid the commotion, a more fundamental question has come to the forefront: How far will this price war go?

From the perspective of industry evolution, video large models appear to be replicating the price curve traversed by large language models—technological maturation, declining costs, and players using low prices to gain scale. However, there is a decisive difference: higher inference costs mean that a purely price-driven strategy cannot be pursued to its full extent.

Because the true floor price is never determined by competitive willingness, but rather supported by mining cost.

Taking an open-source video model with 13 billion parameters as an example, the inference phase demands extremely high video memory bandwidth and memory capacity, with generating a single video consuming substantial GPU resources. Even after reducing costs by half through model compression and inference architecture optimization, computational power still has a hard lower limit. At $0.5 per second, we are already approaching the cost threshold under current technological conditions; further reductions would either compromise quality or result in losses.

Therefore, what follows isn't about who can drive prices below floor level first, but who can successfully integrate video generation into real-world workflows.

On this point, the positions of several parties are vastly different.

Ji Meng’s moat is the Douyin and CapCut ecosystem. Ji Meng AI has surpassed 50 million monthly active users, with its core advantage lying in seamless data integration with CapCut—logging into the same Douyin account activates features like automatic asset import/export, one-tap navigation, and automatic subtitle track loading. CapCut bets on the existing workflow: users open CapCut specifically to edit videos, modify assets, and post to Douyin; Seedance is simply an extra hand in this established process. ByteDance’s goal is not to pull users to a new platform, but to embed model capabilities directly into the existing actions of creators, advertisers, and merchants. Creators don’t need to migrate—the toolchain is natively connected—and this is precisely the foundation of Ji Meng’s confidence in its three consecutive upgrades.

Keling’s moat lies in Kuaishou’s content community and creator ecosystem. The Keling 3.0 series, built on an All-in-One philosophy, integrates video understanding, generation, and editing into a seamless AI workflow. Unlike most tools on the market that still remain in the early “pull-and-generate” stage, Keling has already served over 60 million creators worldwide and generated more than 600 million videos. It is a daily tool for short-form video creators, naturally embedded within the Kuaishou ecosystem—the very territory it defends.

Alibaba's strategy is "openness + price reduction." HappyHorse is deployed on Alibaba Cloud’s Bailian platform and offers open API access, enabling end-to-end integrated content creation workflows—from text-to-video and image-to-video to reference-based video and video editing. Rather than building a closed ecosystem, it delivers its capabilities outward for third-party platforms to integrate—Wanxing JuChang is the first example. This ecosystem aggregation approach aligns perfectly with ATH’s overarching logic of “creating Token, delivering Token, and applying Token”: the model serves as the entry point, while the invoked Token is the ultimate goal.

Seedance Mini targets developers. Volcengine plans to soon launch its API services for high-frequency, scalable scenarios such as e-commerce content creation, bulk generation of marketing materials, and UGC content production. ByteDance is betting on something else here: that video generation capabilities will be integrated into more applications and SaaS products—enabling developers to invoke them, rather than requiring end users to use them directly. Its goal is to become an invisible yet ubiquitous underlying infrastructure behind countless products.

It is evident that the video large model market appears to be splitting into two tiers.

The upper layer is the value-premium market. Cinematic creators, brand advertisers, and professional studios are willing to pay for high-quality output. They seek premium results like Seedance 2.0 Standard or Ke Ling 3.0, are insensitive to price, but extremely sensitive to quality. Jiemeng’s pricing strategy has a solid foothold in this market—while users may complain about the high cost, as long as the quality remains领先, price increases won’t drive away core users.

The lower tier is a cost-sensitive market. E-commerce short-video teams, localized advertising agencies, and short-form drama production companies need video素材 that is cheap and stable enough. They don’t seek cinematic quality—only high-volume output and rapid iteration. For these users, Seedance Mini at $0.5 per second and the continuously declining prices of HappyHorse are the optimal solutions.

The current price war is merely each player vying for a ticket to their respective market segment. When prices hit the cost floor and subsidies can no longer be sustained, competition will revert to its essence: for consumers, it’s about creative efficiency within an ecosystem闭环; for businesses, it’s about the depth and breadth of integration into workflows.