Walk into a camera-rental house a year after a new flagship body has arrived and an old rule of technology becomes visible on the shelves. Yesterday’s top camera is still capable of making beautiful images, but its rental rate has fallen. The manufacturer has recovered part of its development cost, competitors have caught up, and the market now regards the same machine as a previous-generation product. Creative technology depreciates.
AI video appears to break that rule because an AI model is not merely a camera. It is a camera, a render farm and a film laboratory bundled into a metered service. The customer never buys the instrument; the customer pays for each act of computation. Every generated frame must be manufactured from noise, whether the model was released yesterday or last year. When a new camera arrives, the old camera remains on the shelf.
When a new video model arrives, the old model still has to perform billions of calculations every time somebody asks it to make a shot.
That makes the arrival of ByteDance’s Seedance 2.5 more than another entry in the increasingly crowded model league table. It raises an uncomfortable economic question. If each substantial improvement demands more training compute, larger datasets and more expensive inference, do we eventually reach a model that is technically impressive but economically final?
Is Seedance 2.5 close to the point beyond which creators—or even the companies building these systems—can no longer afford the next step?
The short answer is probably no. The more useful answer is that the question exposes a genuine fracture in the economics of AI: the frontier is becoming dramatically more expensive to create at exactly the same time that yesterday’s capabilities are becoming dramatically cheaper to use.
What Seedance 2.5 Is Actually Adding to Character AI
The pace of ByteDance’s recent development helps explain why the concern feels plausible. Seedance 2.0 was officially launched in February 2026. It could accept text, images, video and audio as references, generate synchronized audio and video, and produce 15-second multi-shot outputs. According to ByteDance’s release material, a user could provide as many as nine images, three video clips and three audio clips in a single request.
Seedance 2.5 followed only months later. It doubles the maximum single generation to 30 seconds and expands the reference allowance to 30 images, ten video clips and ten audio clips. It offers multi-round extension, timestamp-level editing, improved green-screen work, more controlled camera perspectives and stronger continuity across scene changes.
ByteDance is positioning it not simply as a better clip generator but as a system that can interpret a package of creative materials and carry an idea further into production.
In this context, character AI is not merely a conversational interface. It is the model’s ability to preserve identity, performance, wardrobe and spatial continuity as a character moves through a shot or across multiple scenes.
For filmmakers, that distinction matters. A longer clip is not automatically a better shot, and a model that invents more of the sequence can also invent more mistakes. Yet the move from a prompt and a start frame toward references, blocking, editing instructions and temporal control represents genuine progress. It brings the model closer to the way directors already work: through accumulated decisions rather than a single verbal description.
It also makes the computation more demanding. Video already combines spatial complexity with time. Longer duration means more frames; higher resolution means more information within every frame; synchronized sound adds another generated stream; reference videos must be interpreted before the output is produced. In current video diffusion architectures, a five-second 720p clip can unfold into more than 100,000 internal tokens.
Researchers working on efficient video diffusion describe attention—the mechanism that relates those tokens to one another—as the principal scaling bottleneck.
From that perspective, the fear of a cost ceiling seems rational. Every model announcement promises more: longer shots, higher resolution, stronger physics, native sound, more references, better editing and fewer continuity errors. Each promise appears to add another load to the machine.
The Two Cost Curves of Artificial Intelligence
The first curve is the one attracting most of the alarm. Frontier AI development is becoming extraordinarily capital-intensive. Epoch AI’s analysis estimates that training compute for frontier language models has grown roughly fivefold per year since 2020, while training cost has risen about 3.5 times per year. Power requirements have doubled annually. A typical AI data center with one gigawatt of IT capacity now represents approximately $38 billion in upfront capital.
Video-model developers disclose far less than would be required to make an equivalent calculation, so language-model figures should not be casually presented as Seedance’s costs. They do, however, show the direction of foundation-model development. The most capable systems increasingly require infrastructure that only a handful of companies or state-backed organizations can assemble. Five hyperscalers already control more than two-thirds of global AI compute.
If this were the only curve, the conclusion would be grim. Frontier models would become ever more expensive; API providers would either pass the cost to customers or sell generations below their full cost; independent labs would disappear; and creative access would depend on the willingness of ByteDance, Google, OpenAI and a few other companies to subsidize it.
But a second curve is moving in the opposite direction. The cost of obtaining a fixed level of AI capability is falling extremely quickly. Stanford’s 2025 AI Index found that the inference cost of a system performing at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. Hardware costs declined around 30 percent annually, while energy efficiency improved around 40 percent per year.
More recent research from MIT estimated that the price of achieving a fixed level of performance was falling five- to tenfold per year, with algorithmic improvements accounting for roughly a threefold annual efficiency gain.
This is why a newer model can be more capable and cheaper. GPT-4o launched at half the price and twice the speed of GPT-4 Turbo. Google reduced Veo 3 from its original $0.75 per second to $0.40, while introducing Veo 3 Fast at $0.15 per second.
New hardware, quantization, distillation, sparse computation, caching, better batching and improved serving software continually change the amount of useful work that can be extracted from a dollar of compute.
The old model’s arithmetic does not magically disappear, but neither is it frozen in the data center where it was born. The weights can be served on newer chips and through better software. More often, providers do not reduce the price of an old product indefinitely. They retire it or replace it with a smaller, faster model that delivers comparable results more efficiently. Yesterday’s intelligence becomes a feature inside tomorrow’s budget model.
What a Second of AI Video Really Costs
Seedance pricing itself reveals why simple generational comparisons are misleading. On Runway’s API, which offers several third-party models under one billing system, Seedance 2.0 currently costs $0.36 per output second at 720p. Seedance 2.5 costs $0.30 per second at the same resolution, although input and reference video incur additional charges. At 1080p, the relationship reverses: Seedance 2.0 costs $0.40 per second and Seedance 2.5 costs $0.68.
The full pricing table contains similarly wide differences between standard, fast and lightweight variants of other models.
BytePlus makes the variability even more explicit. Its five-million-token Seedance 2.5 package is advertised as producing approximately 120 to 500 seconds of 480p video. That is more than a fourfold range from the same token allowance, depending on the generation mode and inputs. There is no honest single answer to the question, “What does Seedance 2.5 cost per second?” without specifying resolution, duration, references, provider and workflow.
For a working creative, however, even that calculation is incomplete. The useful metric is not cost per generated second but cost per accepted second.
Suppose an older model costs 30 cents per second but requires eight attempts before it produces a shot with acceptable anatomy, performance and continuity. A newer model at 60 cents per second needs only three. The supposedly expensive model has already won before anyone counts the hours spent masking errors, rebuilding references or persuading a client that the actor’s coat changed color for artistic reasons.
AI video pricing rarely includes failure yield, yet failure is where a large share of production budgets goes. As Ciaro Pro’s guide to budgeting an AI film argues, the visible generation fee is only one part of the production cost; planning, continuity, iteration, editorial work and finishing determine whether those generated seconds become a film.
A model that follows the storyboard, respects the reference performance and preserves the set across a camera move may be economically superior even at twice the list price. Reliable character AI can therefore justify a higher generation price when it reduces rejected takes and continuity repairs. Conversely, a spectacular model can be poor value when its best qualities are irrelevant to the shot. A locked-off insert does not require the industry’s most expensive reasoning about physics and dialogue.
This changes the role of production software. The valuable layer may not be a permanent allegiance to whichever model currently leads the rankings, but an AI movie-making system that keeps the script, boards, characters, references, versions and edit connected while different generation engines are used for different shots. As models become more specialized and their price curves diverge, routing becomes a creative and financial decision.
That routing also demands category discipline: poly ai and polybuzz ai serve conversational use cases, magicschool ai focuses on education, humanizers ai, humanize ai and ai humanizer concern text transformation, while google ai studio and an ai mode support broader creation or search workflows—not shot-level video economics.
The Ceiling May Change the Industry Without Stopping Progress
None of this eliminates the frontier-cost problem. It changes its likely consequence. We are less likely to encounter a final video model than a market in which frontier development is concentrated among companies able to justify it through several businesses at once.
ByteDance does not need Seedance to behave like a standalone camera company. The model can support consumer creation products, cloud services, advertising tools and internal media workflows. ByteDance also says Seedance 2.5 is being applied to synthetic training data, industrial simulations, robotics and autonomous-driving scenarios. The same research investment can therefore produce value across markets that no subscription from independent filmmakers could support by itself.
Google can make a similar calculation across cloud infrastructure, advertising, YouTube and Workspace. Meta can fund models through an advertising business. A specialized model company without those surrounding economics faces a harder problem. Even when inference on a paid API request carries a healthy margin, recovering billions in research, data acquisition and infrastructure is a separate matter.
The danger is therefore not that video models suddenly stop improving. It is that the frontier becomes inseparable from a small number of corporate ecosystems. Creative access could remain affordable while control over the underlying technology narrows. Prices may be low because efficiency improved, because a platform is acquiring customers, because another division benefits from the model, or because a provider is engaged in a price war.
Those situations look identical on an API invoice but create very different long-term dependencies.
There is also a technical reason not to declare a final model. The pressure created by expensive scaling is producing more efficient architectures. Video Sparse Attention researchers have reported a 2.53-fold reduction in training computation without a corresponding increase in diffusion loss. On a 14-billion-parameter video model, their method reduced generation time from 1,274 seconds to 576 seconds.
Combining sparse attention with distillation produced a reported 50.9-fold speedup on a smaller model while maintaining quality.
Those figures come from research systems, not evidence that the next commercial model will be 50 times cheaper. They nevertheless demonstrate that progress is not limited to buying more GPUs. A model can improve because its developers discover which calculations were unnecessary.
So, Is Seedance 2.5 the Last Model We Can Afford?
Almost certainly not. Seedance 2.5 may even become much cheaper to use while a more capable successor occupies the premium tier. That has been the dominant pattern across language, image and video generation: the frontier moves upward while its previous capabilities diffuse into faster and less expensive systems.
The serious limit lies elsewhere. Brute-force scale cannot remain the only reliable route to improvement. If every gain in coherence or physical accuracy requires a geometrically larger training run and more computation for every generated frame, the number of organizations capable of competing will continue to shrink. The next decisive breakthroughs will have to improve the relationship between computation and usable creative control, not merely the benchmark score.
For filmmakers evaluating character AI in video, this suggests a less glamorous but more durable way to evaluate progress. Ask how many generated seconds survive the edit. Ask whether references remain stable, whether the model can revise one moment without destroying the rest, and whether a cheaper engine can handle the shot. The winning system will not necessarily generate the most astonishing clip in isolation.
It will waste the least money between an intention and a finished sequence.
Seedance 2.5 is not the end of AI video. It may, however, mark the point at which the economics become impossible to ignore. The future of generative filmmaking will be determined not only by what a model can imagine, but by how efficiently that imagination can be directed, repeated and paid for.
