Inkitt turns its AI video workflow into a product

What Inkitt actually built

Inkitt is a digital publisher and AI-media startup best known for a platform where users publish and distribute ebooks. Its video effort grew from a specific production problem: the company had a large catalog of stories it wanted to adapt for Inkitt Ironblood, its AI-video streaming service focused on action, adventure, science fiction and fantasy.

Traditional production economics could not support that volume. Cinematica began as an internal tool around the beginning of 2026 and had reached its third version by the time VentureBeat saw it.

Inkitt did not try to compete directly with ByteDance’s Seedance, MiniMax and the growing set of video-generation systems. Founder and CEO Ali Albazaz told VentureBeat that the opportunity was to make existing models easier to control across a complete production process.

The demonstrations showed a workflow in which a creator can:

The creator is not locked to one video generator. Cinematica can choose systems based on quality and cost. Inkitt uses Seedance 2.5 from ByteDance for more expensive, higher-quality generation and can offer MiniMax H3 when lower costs matter.

The difference from using a model directly is operational. A user working through a site such as fal.al may have to switch between several menus and assess clips one by one. Cinematica presents those steps as one film-making process, with continuity, production techniques and model selection handled above the individual generators.

Albazaz compared the internal system with a Formula 1 car that Inkitt is now turning into a Honda CR-V for external users. The analogy captures the product challenge: a tool built for expert operators has to become usable by customers who do not share the original team’s knowledge.

The framework is the accumulated workflow

Inkitt’s approach starts with people who already understand the work. Around 25 employees were working on Ironblood and Cinematica at the time of the interview, including approximately 15 video producers, and the company was still hiring.

Those producers are not merely users of the system. They help develop it.

Inkitt holds daily meetings where the production team discusses techniques discovered the previous day and decides how to add them to Cinematica. If Jessica finds a useful technique, the team turns it into a Cinematica feature. The same process applies to discoveries made by Jose and other employees.

That creates a feedback loop:

The technique can be specific. If a character’s skin looks waxy, Cinematica can apply methods producers use to correct that defect. If a scene needs a particular cinematic effect, it can suggest camera or lens settings.

Some problems require more elaborate fixes. Albazaz said video generators can change faces or character positions when several people appear in a moving shot. One way to address that is to arrange the characters with a 3D- or CAD-like layout that helps the generator preserve their spatial relationships.

This is a different automation strategy from trying to remove the specialist. Inkitt places specialists inside the improvement cycle and converts repeated decisions into software. The valuable enterprise asset is therefore not just proprietary data. It can also be the implicit knowledge employees use when models fail.

The framework also draws on knowledge from outside Inkitt. Producers follow techniques shared by the AI-video community, particularly on X, attend creator meetups and combine public methods with their own production experience.

That knowledge does not need to be invented internally to become useful. The company’s advantage may come from collecting, evaluating, standardizing and continuously updating methods, then connecting them to its own workflow.

The architecture protects Inkitt from model churn

Cinematica’s main architectural choice is that it is not a foundation model. Its value is intended to sit above the generators:

For an enterprise, that separation reduces dependence on whichever supplier is strongest today. If another model becomes better or cheaper, it can be substituted while the company keeps its workflows, rules and expertise.

Salesforce is promoting a related idea with an organizational control layer that works across different AI models and agent platforms. Cinematica is much narrower, but the architectural principle is similar: the company owns the workflow and orchestration layer even when another vendor owns the underlying model.

I think this is the more interesting part of Inkitt’s launch. The public beta is not primarily a bet that Inkitt can beat dedicated model developers at video generation. It is a bet that production coordination is difficult enough, and valuable enough, to justify a separate product.

That bet is reflected in the pricing. Inkitt estimates a finished, polished minute of content at $75–$500, depending on the number of generation attempts and the model selected.

Movie Creator’s credit packages, observed by VentureBeat, were:

That works out to approximately $20.60–$21.20 per calculated minute, with almost no discount at larger package sizes.

For comparison, BytePlus prices Seedance 2.5 at about $27.72 per minute at 720p and about $12.34 at 480p. Google’s Veo 3.1 costs $24 per minute for standard 720p or 1080p generation with sound, $6 for Veo 3.1 Fast at 720p and $3 for Veo 3.1 Lite at 720p. MiniMax charges $4.80 per minute for H3 at 768p or $7.80 at 2K.

Movie Creator is therefore cheaper than direct 720p access to Seedance 2.5, but more expensive than 480p Seedance and the cheaper Veo and MiniMax options. The comparison is not like-for-like: those APIs sell raw model inference, while Inkitt’s price includes orchestration, resource management, continuity, model selection and accumulated production techniques.

Still, the gap is material. My guess is that Cinematica will only justify it for customers who value fewer failed attempts and less manual coordination, not simply more generated footage.

Inkitt is selling production capacity, not just video

Movie Creator became available through Inkitt’s website on September 24, 2026. There is no Cinematica API at launch, so production remains inside the site. Inkitt has an MCP server, but Albazaz said it is not yet connected to Cinematica; integration is planned for later.

That makes the first version a creative production tool rather than infrastructure that can plug directly into a large enterprise marketing stack.

Inkitt’s wider strategy explains why this workflow exists. The company began as a platform where authors published fiction and Inkitt analyzed reader behavior to identify promising stories.

In 2017, it raised $3.9 million around its “reader-powered publishing” model. By 2019, it had 1.6 million readers, 110,000 authors and approximately 350,000 uploaded stories, according to TechCrunch. It later expanded successful works through the paid fiction service Inkitt Galatea, then raised $59 million in a 2021 Series B round while broadening its plans toward multimedia entertainment.

In 2024, Khosla Ventures led a $37 million Series C round. Inkitt told TechCrunch then that it had reached 33 million users and planned to expand into audiobooks, games and AI-video creation. Total disclosed funding reached $117 million.

The company subsequently launched the short AI-drama platform Inkitt CandyJar and Inkitt Ironblood. Movie Creator now exposes part of that internal video operation to outside users.

One early commercial use is branded short dramas. Albazaz said companies had approached Inkitt about integrating products and services into serialized stories, including an unnamed apartment-search company working on a story about a couple looking for housing. A cruise company had also shown interest.

The format is already being tested by brands including:

Fast Company reported that Crocs’ “Charmed to Meet You,” centered on Jibbitz charms, collected nearly 10 million views. JCPenney’s partnership with TelevisaUnivision generated 16 million impressions and 5.6 million video views, according to Marketing Dive.

The business outcome remains unsettled. Marketers are still working out how reliably interest in short dramas turns into site visits and sales.

The Wall Street Journal reported that branded short dramas in the United States typically cost about $200,000–$450,000 to produce. At Inkitt’s stated $75–$500 per finished minute, 60 minutes from Movie Creator would cost $4,500–$30,000 before additional human work and external post-production. Inkitt’s output can still be moved into tools such as Premiere for editing, sound effects and other finishing work, so the figure is a generation and AI-production cost, not evidence that the entire traditional budget has been replaced.

The missing evidence is the product’s real test

Inkitt has shown a production workflow, but not yet measured Cinematica’s contribution separately from the underlying models.

The company did not provide VentureBeat with benchmarks comparing Cinematica with direct use of Seedance, MiniMax or another video model. It has not published:

That omission matters because the central claim is about the framework, not simply the quality of the models it selects. Xiaomi’s HarnessX, for example, improved results in 14 of 15 model-and-benchmark combinations, with an average absolute gain of 14.5% without replacing the underlying model.

In its demonstration, Albazaz showed a two-minute clip that he said an Inkitt producer assembled in approximately 25 minutes before final post-production. That is useful evidence of what the team can do, but it does not measure the average production-time reduction or establish how the result compares with direct model access.

What I’d want to know is whether Cinematica consistently lowers the number of failed generations enough to offset its price premium. Without that measurement, the framework’s value is plausible but unproven.

Inkitt’s experience does not mean every company should build its own video platform. It suggests that a useful industry framework may already have the same basic ingredients:

If Inkitt can quantify the benefit, the next layer of enterprise AI competition may belong less to companies with exclusive access to the strongest model and more to companies that can turn employee know-how into durable software above it.