The most exciting area of development in enterprise AI right now isn't models, nor is it even agents: it's the harness.
Companies including Salesforce, DeepSeek Xiaomi, Google are increasingly offering new, custom harnesses designed to enhance underlying AI models and agent performance, and improve their reliability on specific, repeatable, yet open-ended tasks. The harnesses are meant to achieve this by providing the AI model with a layer of software, tools, prompts, context, execution environments and operating instructions dedicated to accomplishing the sought-after goals.
But the natural question arises: if these firms can do it, should your company also be building its own dedicated harness to accomplish the specific tasks applicable to your business? And can your own harness be productized? One example of a firm that did it recently is the digital publishing and AI media startup Inkitt, best known for allowing users to publish and distribute e-books to its community. Today, it launches Inkitt Movie Creator, a public beta designed to turn text-based scripts into AI-generated films in minutes.
Inkitt Movie Creator is powered entirely by Cinematica, an internal harness created by Inkitt's in-house AI video production experts that can orchestrate different image, video and language models while encoding real-world filmmaking techniques.
Instead of building another foundation video model to compete directly with ByteDance’s Seedance, MiniMax and the expanding field of generative-video systems, Inkitt founder and CEO Ali Albazaz told VentureBeat in an exclusive interview that his company is betting the larger opportunity lies in making those models easier to control across an actual production.
“There’s a lot of choice there, and we don’t want to compete there at all,” Albazaz said. “Cinematica — what makes Cinematica special is the harness.”
Research has increasingly shown why harness engineering matters. Xiaomi’s HarnessX improved performance across 14 of 15 model-benchmark combinations in testing, producing an average absolute gain of 14.5% without simply swapping in a stronger foundation model.
Separate real-world testing covered by VentureBeat has shown the same model producing substantially different results depending on its harness, tool configuration and execution stack.
Cinematica represents a particularly tangible attempt to apply that same principle outside coding and enterprise agents: rather than train a better movie model, build a better system for directing the movie models that already exist. However, the company declined to release specific benchmarks on its own harness at launch time, instead providing VentureBeat with demoes and walkthroughs showing how it sped up production and delivered editable results in moments, creating an efficient workflow for AI video production that can be used in advertising or promotional materials.
In my view, the demoes did show an improvement in ease of user experience over using AI video models directly on the vendor websites, or even through a multi-model aggregation service such as fal.al, where the user has to click through multiple dropdowns and evaluate AI generated video clips piecemeal, rather than as part of a unified, single movie.
And most importantly, for enterprises wondering whether they should build a custom harness of their own, Inkitt’s development process offers at least five useful lessons.
Lesson 1: Start with a real internal workflow, not a generic harness idea
Cinematica did not begin as an abstract platform project in search of a use case.
According to Albazaz, it originated around the beginning of 2026 because Inkitt already possessed a growing catalog of stories that it wanted to turn into films for Inkitt Ironblood, it AI video streaming service for action, adventure, sci-fi and fantasy films based on successful Inkitt e-book titles. As of now, all the content on the site was made with Cinematica, a spokesperson told VentureBeat.
Conventional production economics made adapting them at Inkitt’s desired volume impractical, while increasingly capable AI video models offered another route.
“We started to create Cinematica version one,” Albazaz said. “It started as an internal tool for us.” The system shown to VentureBeat is now version three.
In other words, Inkitt did not start with the proposition that it needed “an AI harness.” It started with a recurring business problem: how do we use rapidly improving video models to produce many longer-form movies without rebuilding the production workflow around every new model?
Cinematica emerged as the answer. Albazaz used an automotive analogy to describe the subsequent transition from internal infrastructure to public product.
“We created this Formula One car, if you will, to help us create these movies internally, and now we're turning it into a Honda CR-V so that externals can start using it,” he said.
That progression offers a potentially useful template for other enterprises: start with a real, repeated and expensive internal workflow, solve it for your own employees first, and only then consider whether the resulting harness should become something broader.
Lesson 2: Put domain experts inside the harness-development loop
The humans doing the work are central to Inkitt’s approach. Albazaz said roughly 25 people were working across Ironblood and Cinematica at the time of the interview, including about 15 video producers, and that Inkitt was continuing to hire more.
Those employees are not simply end users of Cinematica. They effectively function as an ongoing R&D organization for the harness.
Inkitt holds a daily meeting where the production team discusses techniques discovered during the previous day's work and considers how to incorporate them into Cinematica, Albazaz said.
His description of the process was strikingly straightforward.
“'Jessica, you learned something new today. Let’s productize it into Cinematica so that the rest of the team can benefit from your learning',” he said, giving a hypothetical example. “'Jose, you just learned something today. We're going to productize it, put it into Cinematica, so that the rest of the team is upgraded.'”
The result is that "everybody’s learnings on a daily basis are basically fed into Cinematica," Albazaz told me.
That is a potentially important inversion of the typical automation narrative. Instead of asking an AI system to eliminate a domain expert's job and somehow acquire their expertise independently, Inkitt puts those experts inside the optimization loop and tries to turn their repeated decisions into reusable software.
For an enterprise, the valuable asset may therefore be not only proprietary data but the tacit operational knowledge held by people who already understand the workflow better than the model does.
Lesson 3: Turn recurring failures and expert fixes into reusable system knowledge
That human input becomes most valuable when it is attached to concrete failure modes.
Albazaz described Cinematica as capturing techniques for problems that repeatedly emerge during AI-video production.
If an AI-generated character's skin looks waxy, the harness can encode techniques producers have found useful for correcting it. If a particular scene calls for a certain cinematic effect, it can help recommend camera or lens setups.
More complicated scenes can demand stranger interventions.
Albazaz described video generators sometimes swapping characters’ faces or positions when multiple people appear in a moving shot. One workaround involves arranging the characters with a 3D or CAD-style reference that helps the underlying generator preserve their spatial relationships.
“The harness is being trained on, like, ‘Hey, whenever you are facing this issue, there is this technique,’” Albazaz explained. "Theoretically, you could go on Twitter and Google and whatever, and try to, every time you face an issue, Google it yourself. But we have basically put all the tricks into one place.”
In effect, Inkitt is attempting to turn recurring troubleshooting into institutional memory.
A coding team might encode repository conventions, debugging sequences or preferred tool-use patterns. A security organization could capture the investigative steps senior analysts repeatedly use for particular classes of alerts. A finance team could encode how experienced employees reconcile certain anomalies before escalating them.
Inkitt’s domain happens to be filmmaking, so the accumulated knowledge concerns characters, lenses, continuity, prompting and visual artifacts.
But the underlying pattern is portable: find where the model repeatedly fails, observe what an expert does next, and make that intervention available the next time the system encounters the same class of problem.
Lesson 4: Combine proprietary expertise with useful knowledge from outside the company
Inkitt also does not claim that all of Cinematica’s production knowledge was invented internally.
Albazaz said its producers monitor techniques shared publicly by the wider AI-video community, particularly on X, while also developing their own approaches through production work.
“It is a combination of us learning from the community, but then also from the producers themselves,” he said.
Inkitt encourages its producers to spend time keeping up with newly discovered techniques. Albazaz also cited conversations at gatherings of AI-video creators as another source of practical knowledge that can make its way back into the system.
The competitive advantage, then, does not necessarily have to come from inventing every underlying technique.
It can come from collecting, evaluating, standardizing and continuously updating the techniques relevant to one particular job, then combining generally available knowledge with internal methods and workflow data that competitors do not possess.
That process also keeps the harness from becoming static.
During VentureBeat’s later demo, Inkitt executives emphasized that Movie Creator users are intended to benefit as its internal production team discovers new techniques rather than waiting for occasional major product versions.
“It’s also not stagnant,” one Inkitt executive said. “All of the prompts that they’re doing, all of the feedback, the smarter it gets, that’s constantly being relayed then to the users on Inkitt.”
Lesson 5: Keep the harness separate from the underlying models
Inkitt’s final architectural choice may be the most directly applicable to enterprises: the harness is not the foundation model.
During a demo for VentureBeat, Cinematica showed a workflow in which a filmmaker can establish characters, locations and props; generate several versions of individual scenes; select preferred generations; and ask the system conversationally to modify shots that are close but wrong.
But creators are not locked into one underlying video generator.
Albazaz said Cinematica can select among different systems according to factors such as quality and cost, citing ByteDance’s Seedance 2.5 for more expensive high-quality generation and MiniMax H3 when users want to keep costs lower.
“If you really want to go high quality, you go with Seedance 2.5, but it becomes expensive,” Albazaz said. “If you want to go high quality really cheap, you go probably with MiniMax H3, and we allow the user to make those selections.”
Inkitt executives described an even wider stack in the later demonstration, combining different providers for image generation, video and conversational interaction.
“We’re not bound to one model,” Albazaz said. “We will give you the best tools.”
That separation matters because the intended value of Cinematica is not merely access to a particular video generator. It is the production logic that persists above those models.
“How do you keep the characters consistent over an hour long? How do you keep the same voices? How do you keep the locations the same?” Albazaz said. “What lenses do you select when you’re shooting it? All these kinds of tricks that we have learned over the months and months of playing around with these different models allow us to create very cinematic movies.”
For an enterprise, that model-agnostic approach could help prevent the business process itself from becoming tightly coupled to whichever AI provider happens to be strongest today.
If another model becomes better or cheaper tomorrow, the model can theoretically change while the organization's accumulated workflow, rules and expertise remain in the harness.
That principle is especially visible in AI video, where model capabilities and pricing are changing quickly, but it increasingly appears across enterprise AI as well.
Salesforce, for example, is pushing the idea toward an enterprise-wide governance layer that works across models and agent platforms rather than assuming companies will standardize on one provider.
Cinematica is considerably narrower, but the architectural instinct is similar: own the workflow and orchestration layer even if somebody else owns the foundation model.
Pricing and availability
Inkitt Movie Creator is available starting today, September 24, 2026 through the company's website. There is also no Cinematica API at launch, so all production will for now, occur on Inkitt's website. While it operates an MCP server, Albazaz said it was not yet connected with Cinematica — an integration is planned.
Those omissions may make Cinematica more useful initially as a creative-production tool than as infrastructure ready to plug directly into a large enterprise marketing stack.
Inkitt estimates that Movie Creator productions will cost “$75 to $500 per minute of final, polished content,” according to a fact sheet provided to VentureBeat, with the wide range attributable to how many generation attempts a creator makes before accepting a scene, and the underlying AI video creation model used, which the user may select. Inkitt serves the models up through Movie Creator and charges an additional premium for the use of its interface, though the company declined to specify how much or if it was variable.
When VentureBeat tested the new creator, we received a pricing screen for bundles of credits beginning at 50,000 credits for $10.61, which Inkitt estimates is enough for roughly 30 seconds of movie generation. Additional packages include:
100,000 credits for $20.91 (about one minute)
200,000 for $41.51 (two minutes)
500,000 for $103.30 (five minutes)
1 million for $206.29 (10 minutes)
2.5 million for $515.25 (25 minutes)
That works out to a remarkably consistent roughly $20.60–$21.20 per estimated minute of generated video, with almost no volume discount as users move into the larger credit bundles.
That pricing is about in the middle of other AI video offerings, including first-party AI video models. Under BytePlus’s first-party Seedance 2.5 pricing, 720p generation works out to about $27.72 per output minute and 480p to about $12.34 per minute, meaning Movie Creator is actually cheaper than direct Seedance 2.5 at 720p but more expensive at 480p.
Google’s latest Veo 3.1 API pricing ranges from $24 per minute for standard 720p/1080p generation with audio to $6 per minute for Veo 3.1 Fast at 720p and just $3 per minute for Veo 3.1 Lite at 720p. MiniMax is cheaper still: its first-party H3 API pricing is $4.80 per minute at 768p or $7.80 per minute at 2K.
Thus, Inkitt’s raw-generation estimate is roughly comparable to premium Veo or Seedance access, but around three to seven times more expensive than directly calling lower-cost Veo variants or MiniMax H3.
That is not an entirely apples-to-apples comparison, however: those APIs sell raw model inference, while Inkitt’s price incorporates access to Cinematica’s higher-level orchestration, asset and continuity management, model selection and accumulated filmmaking techniques — essentially the harness layer this article is examining.
At least against Seedance 2.5 720p and standard Veo 3.1, its apparent per-minute generation cost is competitive; its premium becomes much clearer against lower-cost models such as H3, Veo Fast and Veo Lite. That actually makes the enterprise question more interesting: is Cinematica’s added workflow intelligence is worth paying several times the lowest available raw inference cost?
Inkitt’s path from data-driven books to AI movies
Cinematica also fits a strategy Inkitt has pursued for more than a decade.
The company began as a platform where authors could publish fiction while Inkitt analyzed reader behavior to identify promising stories. It raised $3.9 million around its “reader-powered” publishing model in 2017. By 2019, Inkitt had 1.6 million readers, 110,000 writers and roughly 350,000 uploaded stories, according to TechCrunch.
It subsequently expanded successful stories through paid fiction service, Inkitt Galatea and raised a $59 million Series B in 2021 as its ambitions widened toward multimedia entertainment.
In 2024, Khosla Ventures led a further $37 million Series C. At the time, Inkitt told TechCrunch it had reached 33 million users and intended to move further into audiobooks, games and eventually AI-generated video. The round brought its disclosed funding to $117 million.
That expansion ultimately led Inkitt to launch an AI-generated short-form "microdrama" platform Inkitt CandyJar and, more recently, Inkitt Ironblood, an AI video publishing and streaming platform dedicated to sci-fi and fantasy films based on hit titles from Inkitt.
Inkitt's Movie Creator now opens some of the tooling behind that video-production effort to outsiders.
How enterprises can make use of Cinematica and Inkitt Movie Creator for their own content and advertising efforts
Inkitt sees one particularly timely enterprise application for Cinematica in branded microdramas.
Albazaz told VentureBeat that companies have already approached Inkitt about integrating products and services into serialized stories. He described one unnamed apartment-search company working with Inkitt on a story about a couple hunting for an apartment, along with interest from a cruise line.
“Microdramas are becoming a form of advertising,” he said. “I think this is going to be a great tool for these brands and their comms departments to go and create microdramas that they can distribute wherever they want.”
The wider advertising industry is already experimenting with the format.
Procter & Gamble, Maybelline, Dr Pepper, Marc Jacobs, JCPenney and Crocs have launched branded microdrama projects. Fast Company reported earlier this year that Crocs’ “Charmed to Meet You,” centered on its Jibbitz charms, drew nearly 10 million views, while a JCPenney partnership with TelevisaUnivision produced 16 million impressions and 5.6 million video views, according to Marketing Dive.
The business case remains unsettled. Marketers are still working out how reliably microdrama engagement translates into concrete measures such as traffic and sales. The Wall Street Journal reports that U.S. branded microdramas typically cost roughly $200,000 to $450,000 to produce. By comparison, Inkitt's stated $75–$500-per-finished-minute range would put 60 minutes of Movie Creator output at roughly $4,500–$30,000, before accounting for any additional human labor or outside post-production.
But Albazaz himself said Cinematica-generated material can still go into tools such as Premiere for editing, sound effects and other post-production. So Inkitt's figure is best understood as a generation and AI-production cost, not yet a demonstrated all-in substitute for a conventional production budget.
That potentially makes serialized branded entertainment, product-centered storylines and narrative advertising easier to experiment with before committing conventional production budgets.
The big unanswered question: what are Cinematica's benchmarks?
There is, however, an important question hanging over all five lessons: does Cinematica actually make the same underlying models perform better, and by how much?
Inkitt has not provided VentureBeat with benchmarks comparing Cinematica against using Seedance, MiniMax or another video model directly.
There is no disclosed blind preference test, character-consistency score, regeneration rate, production-time comparison or measurement showing how much the harness reduces the cost required to reach an acceptable shot.
That omission is particularly notable because other harness projects increasingly attempt to quantify their contribution independently from the foundation model beneath them.
During VentureBeat’s demo, Albazaz instead showed a roughly two-minute sample that he said an Inkitt producer assembled in approximately 25 minutes before final post-production.
“This is literally what you can do in 25 minutes,” he said. “Then you go and edit it, and then you add sound effects throughout post-production. That's kind of the real-life capabilities of this.”
Still, the reduction in average production time hasn't been quantified yet. There are other limitations for enterprises at launch.
The lesson from Inkitt is not necessarily that every enterprise should build its own AI filmmaking platform. It is that other enterprises may already possess the ingredients for a useful domain-specific harness: a repeated workflow, foundation models capable of performing pieces of it, experienced employees who know where those models fail, external knowledge that can supplement their expertise, and a growing collection of tricks used to compensate for those failures.
If that thesis holds, the next stage of enterprise AI competition may depend not only on who has access to the strongest models, but on who is best at turning the hard-won expertise of their own people into the harness wrapped around them.