The 'DeepSeek Moment' for video models is not essentially about low prices, but about offering high quality at affordable prices.

On September 28, Kuaishou's Kling AI 4.0 entered internal testing, extending single-shot native video length to 30 seconds with 4K resolution. Simultaneously, the first 'lightweight' model, Kling 4.0 Flash, was launched, emphasizing fast video generation and lower pricing, though its API pricing has not yet been announced.

On October 8, Shengshu Technology released Vidu Q4 Preview, supporting 15 reference images, three reference audio clips, and 4K direct output, while maintaining its advantage of being 'both good and fast'—they set the initial price as low as 0.09 yuan per second.

Think of AI video as a car race: now that the engines are powerful enough, the focus shifts to which car is more fuel-efficient.

Beyond cost reduction, as production capacity continues to ramp up, AI short dramas are evolving from low-budget experiments to high-quality productions, with rapid growth in new categories like AI animated dramas and short series.

Today, the explosion in content demand has attracted a large number of vendors, forcing the industry to establish 'standards': Are character emotions nuanced enough? Does the narrative engage the audience? Can the overall quality support the final product?

With improved model capabilities, content that was once criticized for its 'AI feel' is now indistinguishable from human-made content:

What once required an animation team several months to handcraft can now be perfectly executed by AI: flying calendars pile up to form a skyscraper, then slowly collapse toward the protagonist... The sense of urgency is palpable, leaving viewers 'stunned and slumped.'

Even humorous content styled after vintage news reports, with noisy visual snow and special effects-style fight scenes, has become a 'comfort zone' for AI videos.

Since MiniMax H3 set a 'rock-bottom price' in July, the trend of affordable pricing has surged. More and more creators now feel confident experimenting extensively with AI video formats for expression. Thus, from Kling to Shengshu, cost optimization has become a key focus in the latest round of model competition.

However, in industrial applications, rather than just competing on low prices, the industry needs to pursue the Pareto frontier—finding an optimal balance between the most powerful and the most affordable options.

Simply put, everyone in the industry is striving for both affordability and quality.

Such a turning point has been experienced in the large model industry before. Early last year, DeepSeek offered near-top-tier capabilities at a fraction of the price, making cutting-edge models accessible to all rather than exclusive to a few players—shifting industry competition from 'who is stronger' to 'who offers better value.' This milestone is known as the 'DeepSeek Moment.'

The 'DeepSeek Moment' for video models is not essentially about low prices, but about offering high quality at affordable prices.

An AI short drama industry insider told Guangzhui Intelligence that they still rank their scripts by quality, using more capable models only for top-tier scripts. However, the cost of such models exceeds their budget. While models need to enhance intelligence, the short drama industry requires affordability to standardize AI drama production.

More noteworthy than price wars are the advancements in model capabilities.

Rather than just improving single-point capabilities, model evolution now leans toward serving film industrialization: content coherence, performance control, cinematography, and the integration of special effects with narrative—models have moved beyond the novelty phase and are now truly integrated into production workflows.

With price and capability battles raging on two fronts, both point to the same conclusion: AI video has entered the era of film industrialization.

The National Day Showdown in Large Models: The DeepSeek Moment Arrives for the Video Industry

Several leading AI video model companies chose to make their moves around National Day. What appears to be a price war is, at its core, a comprehensive competition in performance and pricing.

The most visible result is that new models continue to improve in quality while prices start to decline significantly—a signal that the video industry's DeepSeek Moment has arrived.

Although the full pricing for Kling 4.0 and its Flash version has not been announced and is currently in limited internal testing, the Flash version is positioned as 'fast generation with cost-effective pricing.' This indicates that Kling has also prioritized cost-effectiveness in its new version.

Based on Kling's current Q3 API pricing (720P at ~0.9 yuan/sec, 1080P at ~1.2 yuan/sec, 4K at ~3 yuan/sec), the Flash version will likely push entry-level prices even lower.

Shengshu, on the other hand, has slashed prices across dimensions. Vidu Q4 Preview starts at a promotional price of 0.09 yuan/sec, with official estimates suggesting up to five times more content can be generated within the same budget.

From a pricing perspective, the SaaS side offers a two-month limited-time discount, reducing some specifications to pennies per second; the MaaS side prices image-to-video at ~0.6 yuan/sec for 720P and ~0.75 yuan/sec for 1080P, matching or even undercutting the previous Q3 Pro (0.63 yuan/sec).

ByteDance's Seedance next-gen flagship model, however, has increased prices by at least 50% compared to its predecessor: Seedance 2.0's 720P reference-free generation costs 46 yuan per million tokens, while Seedance 2.5 is priced at 70 yuan per million tokens for the same specification.

In contrast, Shengshu's flagship pricing has decreased, showing sincerity.

As vendors unveil their new pricing strategies, the entry price for flagship video models has shifted from the 'yuan' era to the 'cent' era. The industry barrier has been lowered by the vendors themselves.

With Chinese vendors driving down model prices, other industries previously hesitant due to cost are now willing to adopt AI videos.

Market data from Yiwu shows that jewelry vendors using AI-generated videos in multiple styles on overseas media have achieved over 1 million views per video, with inquiry conversion rates increasing by 45%.

Why start a price war now? The answer lies in changes in deliverable quality.

What were AI videos like a year ago? Generating ten videos might yield only one usable result, with 'lottery-style' creation being the norm—some companies even hired dedicated 'lottery operators' to break scripts into 15-second shots, repeatedly generating and screening content while adjusting prompts based on visual performance to achieve a usable version.

At that point, price reductions were meaningless. No matter how cheap, if only one out of ten generations was usable, the actual cost for users remained high.

The turning point came earlier this year. Taking Seedance 2.0 as an example, according to BOC International research, its first-try success rate can reach ~90%, compared to the industry average of ~20%. This means 'lottery-style' creation is no longer a barrier.

Feng Ji, producer of *Black Myth: Wukong* and founder of Game Science, stated that the cost of general video production will no longer follow traditional film industry logic but will gradually approach the marginal cost of computing power.

When success rates climb from 20% to 90%, AI video transforms from a 'novelty' into a true 'productivity tool'—toys are judged by fun, tools by cost-effectiveness.

Luo Yihang, CEO of Shengshu Technology, once mentioned that the core issue in AI video generation is consistency. Solving this problem can meet creation demands across various productivity scenarios, enabling efficient, low-cost mass production.

Now that consistency—the biggest bottleneck—has loosened, the industry has naturally entered a cost-competition phase.

This scenario feels familiar in the large model industry.

When leading foundational models are matched by multiple competitors, lightweight models and price wars become inevitable; when video generation quality passes muster, the next battleground becomes affordability and speed.

The price war has begun, but this is only the first layer of the DeepSeek Moment. Deeper changes are occurring at the capability level.

From Generation to Deliverability: AI Video Enters the Industrial Era

The AI video industry has now entered a crucial phase transitioning from generation capability to deliverability. For users, generation is just the baseline; producing content that meets demands is the true competitive edge.

Compared to before, large model companies must now focus on how to make models produce the desired effects for users.

For example, AI acting has been criticized for its 'wooden' performances, erratic camera movements, and artificial-looking special effects—all pointing to one issue: how to better express content for creators.

This year, AI video generation models have begun considering how to integrate into real production workflows, though approaches vary:

Analyzing domestic models, ByteDance's Seedance series leans toward long-form storytelling, character authenticity, and controllability, extending single-generation videos to 30 seconds with timestamp-level editing to streamline post-production; Kuaishou's Kling series focuses on optimizing shot composition and character movements, excelling in complex cinematography like high-speed motion, continuous action, tracking shots, and panoramic views; Shengshu's Vidu series emphasizes 'script-driven creation,' targeting short dramas and animated series with strong narrative demands, focusing on 'directorial capabilities.'

In practical tests, different models' strengths become evident. We tested a 'couple arguing' scene, asking Seedance 2.0 and Vidu Q4 Preview to handle it in an animated short style. Both models followed prompts to complete all dialogue. Seedance 2.0 excelled in details, like the girlfriend's 'bed-pounding' gesture when acting cute, while Vidu Q4 Preview better handled anime-style visuals, with a distinctly Japanese comedic anime flair.

While their approaches differ, all three aim to evolve AI video from mere ' material generation' to delivering 'ready-to-use' finished products.

Taking AI's application in the film industry as a starting point, Guangzhui Intelligence compared the capability evolution of Kling 4.0 and Vidu Q4 Preview based on official demonstrations:

First is 'acting.' Vidu Q3 could already drive emotional changes through shot prompts, but Vidu Q4 Preview further refines expressive granularity: eye movements, pauses, and facial expressions can now amplify character tension as prompted.

For example, in a dialogue test, the protagonist gently turns to look at a falcon perched on their arm without speaking, using only eye contact and a brief pause to convey a 'commanding' effect.

Meanwhile, audio-visual synchronization, now an industry standard, has been integrated into performance systems in both models' updates.

Vidu Q4 Preview impresses with its ability to 'replicate' reference vocal tones. Many AI short dramas suffer from inconsistent character voices—a 'dominant' tone one moment, a cold AI voice the next. Vidu Q4 Preview supports up to three reference audio clips to maintain consistent vocal tones. Feeding it a 'loli' voice, for instance, allows it to replicate that tone.

Kling 4.0, meanwhile, has optimized lip-sync for character dialogue to avoid audience disengagement.

Next is cinematography, where camera movements serve narrative and emotion—e.g., push-ins for tension, orbital shots for atmosphere, tracking shots for immersion.

Here, Kling and Vidu have different strengths: Kling excels in complex motion, particularly fight scenes, with consistent performance in large-scale and continuous action. Its combat scene featuring a cat as a weapon flows seamlessly:

Vidu Q4 Preview has also upgraded its cinematography, stabilizing dynamic shots like fights, sprints, chases, and clashes while supporting FPV first-person perspective creation.

Finally, special effects: Vidu Q4 Preview now handles high-dynamic effects like explosions, fireworks, particles, and energy not just as surface-level displays but as forces impacting the surrounding world—firelight illuminates characters, smoke obscures vision, and shockwaves scatter debris. In a avalanche scene, the result is indistinguishable from human-made content, saving on post-production special effects costs.

Summarizing this round of competition, Kling 4.0 focuses on action and cinematography, while Vidu Q4 emphasizes 'acting' and directorial expressiveness.

Transitioning from generation to standardized deliverability requires more than isolated technical advancements—it demands a holistic industrial capability built around industry needs.

AI Video: Driving Creation into the Industrial Era

In 1895, the Lumière brothers screened *Train Arriving at La Ciotat Station* in Paris, marking the maturation of motion picture technology.

However, it took another half-century for cinema to become a global cultural and commercial phenomenon—after Hollywood established a rigorous studio system, industrialized genre filmmaking, precise division of labor, and a global distribution network.

Artistic prosperity requires standardized production to take root. Without industrial support, early cinema remained a luxury of handcrafted workshops.

Today, the AI video industry stands at a similar crossroads. And it is Chinese model teams driving AI video models toward film industrialization.

In the AI video industry, Chinese models dominate the top tier. On Artificial Analysis's video model leaderboard, Chinese models like Seedance, Vidu, and Kling frequently rank highly. Today, most leading models come from Chinese companies.

Not only do they lead in performance, but in terms of pricing tiers, the API prices of leading overseas models generally remain higher than those of domestic models.

Comparing the prices of leading models, Google's Veo 3.1 Fast costs approximately $0.15 per second with audio, while the full version costs around $0.4 per second. Even the lightweight Veo 3.1 Lite (720P, $0.05 per second) released at the end of March can only match the pricing of domestic mid-to-high-end models, yet it shows significant gaps in image quality and capabilities.

However, starting in the second half of the year, lightweight models are becoming the main drivers of volume in terms of video model consumption structure. Using OpenRouter's token consumption rankings as a reference, models like Veo 3.1 Lite and Seedance 2.0 Fast are seeing higher consumption than other mainstream models.

Fast iteration, deep scene (Note: ' scene ' is kept as-is due to its contextual specificity; if it refers to 'scenarios,' it should be translated accordingly) penetration, and aggressive pricing—the 'three trump cards' of Chinese models are proving difficult for overseas models to counter.

Zooming out, over the past two years, the domestic AI video industry has completed a full journey from catching up, to keeping pace, to differentiation.

In 2022, the fusion of diffusion models and Transformers just began to push the boundaries of visual generation. Overseas, Stability AI open-sourced Stable Diffusion, spawning a vast ecosystem of fine-tuned models. However, domestic tech giants had not yet prioritized video generation, leaving AI startups to lead the charge—such as Shengshu Technology's U-ViT architecture, which even predates the DiT architecture adopted by Sora.

By 2024, Vidu released the first video model in China that fully benchmarks against Sora, with Kling and Jìmèng entering the field simultaneously, marking the first time domestic models stood on equal footing with OpenAI. Once realism was no longer scarce and flagship models from all players could 'fool the naked eye,' the first tier quickly differentiated—with the focus shifting to cost-effectiveness and differentiated capabilities.

Focusing on the domestic first tier, ByteDance, Kling, and Shengshu represent three distinct approaches:

ByteDance is pursuing a 'dual-track' strategy. On one hand, it leverages Seedance's realism in human representation to provide a model foundation for C-end creation, reaching massive users through platforms like Doubao and Jìmèng. On the other hand, it explores professional B-end solutions, such as white-model tools for the film and television industry, moving toward industrial-grade production tools.

Kling is pursuing a cinematic quality route. The Kling series continues to invest in image details, realism, and overall texture, with settings like 4K 60fps and 21:9 ultra-wide aspect ratios, aiming to directly compete with professional film and television production tools.

Vidu is pursuing a route of expressiveness and cost-effectiveness. Compared to the other two, Vidu's advantages lie more in functional innovation and efficiency optimization. For example, its pioneering 'reference-based video generation' and multi-subject consistency later became industry-standard capabilities. Meanwhile, from Q2 to Q4, Vidu continuously compressed generation efficiency—reducing generation time to under 10 seconds in Q2, achieving 16-second synchronized audio-visual output in Q3, and lowering the price of its flagship model to as low as ¥0.09 per second in Q4, making it accessible to all.

In the industry, as expressiveness becomes affordable, tangible changes are already occurring in content production:

For some industries, as costs decrease, users are shifting from fearing 'lottery-like' outcomes to running multiple tests, enabling good ideas to scale in the market.

Take advertising as an example: previously, a single SKU might only have one or two videos due to high production costs, with success left to chance. Now, with lower costs, ten or eight versions can be produced for A/B testing, and the highest-converting one is selected.

Focusing on creators, the lowered barriers and improved quality of AI video enable individual creators to produce works comparable to those of teams, making 'creation for all' more than just a slogan.

The AI short film *The Laid-Off Girl* is a prime example. Created by individual creator Feifeifei, it surpassed 200 million views on Douyin in July after its June release, with Season 2 launching in August. The total views across all platforms for both seasons are estimated to have exceeded 500 million.

Industry practitioners told Guangzhui Intelligence that by using more cost-effective models, they can now control the cost of a single AI short film at around ¥8,000. Compared to live-action short dramas costing hundreds of thousands, production costs have dropped significantly.

Advances in AI video models now enable storytellers to independently produce works.

The 'DeepSeek moment' for video models is the result of surpassing a capability threshold. It means cutting-edge capabilities are no longer exclusive to a few players, with cost-effectiveness becoming the new competitive dimension.

Next, the key to AI video's transition into the industrial production era lies in solving models' expressive capabilities and further integrating AI into real-world industry content production.

The competition boils down to one focal point: whoever first integrates into producers' workflows will secure the next era.