On 30 September 2025, OpenAI launched Sora 2. It shipped with a phone app, a social feed, synchronized sound, and a cameo feature that let you drop your own face into a generated scene. It went off like a firework. The feed filled with viral clips within days, and the app crossed a million downloads in under five days, faster than ChatGPT had managed, even though it was locked to the US and Canada and needed an invite to get in.
Seven months later it was gone. OpenAI announced that the Sora app and website would close on 26 April 2026, with the API following on 24 September 2026. After those dates, the company says, the data tied to your Sora account is permanently deleted.
Sit with that timeline for a second. The most talked about AI video product of the year went from launch to obituary in under a year. Not because it failed. It worked, people loved it, and it still got switched off because OpenAI decided its compute was better spent on coding tools and enterprise products. If you had built a content workflow on Sora, none of your planning, your prompt library, or your team's hard-won instinct for the tool survived that decision. That is the lesson this post is about. In AI video, the leader can disappear, and that turns your choice of model into a strategic bet rather than a settled fact.
How video got generative
Text-to-video is young, and for most of its short life it was a joke before it was a tool.
The widely shared turning point was unflattering. In March 2023, a Reddit user posted a clip of Will Smith eating spaghetti, generated by an early ModelScope model. It was a melting, twitching horror, and it became an informal industry benchmark precisely because it was so bad. If your model could not render a man and a fork without nightmare fuel, it was not ready.
Runway, a New York company founded in 2018, did much of the early commercial work. It helped build the original Stable Diffusion image model, then shipped Gen-1 and Gen-2 in 2023 as some of the first text-to-video models a normal person could actually pay for and use. The clips were short, soft, and physically confused, but they were a product.
Then in February 2024 OpenAI previewed the first Sora and the bar jumped. Suddenly the conversation was not "can it make video" but "how long until this is good enough to use." Google announced its Veo line in May 2024. Kuaishou, a Chinese short-video company, launched Kling the next month. Inside roughly two years the field went from a meme to a crowded market with billions of dollars behind it. The Will Smith test stopped being interesting because every serious model now passes it.
The field as it stands
Here is the working landscape a marketing team faces today. None of this is stable, which is the whole point, but as of mid-2026 these are the names that matter.
Google Veo is the broad all-rounder. It is built by Google DeepMind, and the January 2026 update to Veo 3.1 brought true 4K output at 3840 by 2160, native synchronized audio including dialogue and ambient sound, and vertical 9:16 framing for Shorts and TikTok. Veo's advantage is reach: it is wired into Google's own products and cloud, and it has the compute and the balance sheet of a trillion-dollar company behind it. For most marketing teams it is the safe default.
Runway is the control specialist and the marketer's pick. Its Gen-4 and Gen-4.5 models are built for people who need to direct, not just prompt: camera moves, a motion brush, reference-driven character consistency, and a real editor wrapped around the model. In February 2026 Runway raised 315 million dollars at a 5.3 billion dollar valuation, with Nvidia, AMD, and Adobe among the backers, and its annualized revenue climbed from roughly 70 million dollars at the end of 2024 into the 265 million to 300 million dollar range a year later. Runway is the closest thing the category has to a focused, independent leader, though independent also means it does not have Google's cushion.
Kling, owned by Kuaishou, competes hard on price and cinematic motion. It handles hair, fabric, and liquids well and runs at roughly ten cents a second, which makes it the cheap premium option. It is also commercially serious: Kling's annualized revenue run rate went from about 240 million dollars in December 2025 to roughly 500 million by mid-2026, and Kuaishou said it was weighing a spinoff of the unit at a possible 20 billion dollar valuation. The catch for some buyers is jurisdiction. Kling is a Chinese product, and its terms grant Kuaishou a broad license to host and use what you generate, which matters if your brand or your legal team cares where content is processed.
Behind those three sits a deep bench. ByteDance has Seedance, strong on narrative shots, though it rolled Seedance 2.0 to more than 100 countries and pointedly skipped the United States after pushback over training data. Luma, Pika, and MiniMax's Hailuo all hold real ground on cinematic mood, fast short-form iteration, and low-cost everyday output. The market itself is large and growing fast: one industry estimate put the AI video generator market at about 717 million dollars in 2025, heading toward roughly 847 million in 2026.
The takeaway is not which one wins. It is that there is no single winner, and the practitioner consensus has already settled there. Most production teams in 2026 run two models, not one.
How to actually evaluate a model
"Which is best" is the wrong question because the honest answer is "best at what, for how long, and at what legal risk." Five criteria do the real work.
Quality, for your specific shot. Benchmark scores measure general capability. They do not tell you whether a model can render your product, your spokesperson's face held steady across a cut, or text on a label that does not warp. Test the model on the shots you actually need, not a demo reel. A model that is stunning at sweeping drone footage may be useless at a clean two-second product close-up.
Control. There is a real split between generation and direction. Some models are slot machines: you write a prompt, you pull the lever, you take what comes out. Others give you camera control, reference images, character locking, and an editing timeline. For one-off social clips the slot machine is fine. For brand work that has to match a look across a campaign, you need the controls, and that is Runway's pitch.
Licensing and rights. This is the criterion teams skip, and it is the one that can cost the most. Three questions decide it. First, does your plan grant commercial use? Many tools allow it on paid tiers and forbid it on the free tier. Second, what license are you granting back? Kling's terms, for example, give its parent company a broad, sublicensable right to use your content. Third, and most serious, who carries the legal risk if an output is accused of copying someone's intellectual property? On video, only Adobe Firefly currently offers IP indemnification, where Adobe will defend you and pay damages, and even that sits behind qualifying business plans with a cap and exclusions. Firefly is also the one major generator trained only on licensed and public-domain material. Most other models are trained on scraped web data, and the legal exposure of that sits with you, the user. That is not a footnote. It is a board-level question for a regulated brand.
Cost, measured properly. Subscription price is the visible number. The hidden one is credits. AI video burns credits fast, the best features sit on higher tiers, and the real cost is per usable second of finished footage, after the failed generations you threw away. A tool that looks cheap per month can be expensive per shot. Price the output, not the plan.
Longevity. This is the one Sora just taught everyone. Before you build a workflow on a model, ask who owns it and why they keep it running. A model inside Google or Adobe is a feature of a profitable company with many revenue lines. A model from a focused independent like Runway lives or dies on that product, which can mean sharper focus or sharper fragility. A model from a venture-funded startup running on someone else's compute can be repriced, restricted by region, or shut down on a quarter's notice. None of these is disqualifying. But longevity belongs on the scorecard next to quality, and almost nobody puts it there.
Why a one-model workflow is a trap
Put those five criteria together and the conclusion is hard to avoid. Building your content operation around a single AI video model is a structural risk, not a convenience.
Sora is the obvious proof, but model risk does not only mean a shutdown. It comes in several shapes, and all of them have already happened to someone. A model can be discontinued, like Sora. It can be geo-restricted, like Seedance 2.0 skipping the US. It can be repriced once the venture subsidy ends and the company needs margins. It can change its terms of service so the rights you relied on shift under you. Or it can simply fall behind, because in a field moving this fast last year's leader is this year's also-ran, and Will Smith eating spaghetti is the monument to how quickly the frontier moves.
Every one of those events lands the same way on a single-model team. The prompt library that your team tuned over months stops working, because prompts do not transfer cleanly between models. The visual style your brand became known for is suddenly hard to reproduce. The in-progress campaign stalls. The institutional skill your editors built evaporates. You are not adjusting a workflow. You are rebuilding one, on a deadline, while competitors who hedged keep shipping.
The fix is the same discipline good content teams already apply to publishing platforms, where the smart move is to spread across several channels so no single algorithm change can sink you. Treat AI video models the same way. Run a primary model and at least one credible backup, ideally owned by a different kind of company so they are unlikely to fail for the same reason. Keep your process model-agnostic: storyboard, shot list, brand rules, and review steps should live in your workflow, not inside one vendor's app. Write prompts as documented intent, not as one model's exact syntax, so they can be re-tuned rather than rewritten. And keep the master assets, the source clips and the final renders, exported and stored under your own control, because OpenAI deleting Sora account data is a preview of what platform dependence costs when the platform leaves.
The strategic bet you are really making
Strip away the model names and a content team choosing AI video is making three bets at once, whether it admits it or not.
The first is a capability bet: that this model is good enough at the shots you need today. That bet is the easiest to check and the shortest-lived, because capability is converging and the gaps keep closing.
The second is a continuity bet: that this model, or a clean migration path off it, will still exist when your next campaign ships. Sora is the reminder that this bet is real and that popularity does not secure it. A viral product with a million users still got switched off.
The third is the deepest. It is a bet about where your team's actual value sits. If your advantage is that you know one tool's prompt tricks better than your competitors, that advantage has the lifespan of that tool, and you just watched a seven-month lifespan. If your advantage is editorial judgment, brand taste, story sense, and a production process that can absorb whichever model is best this quarter, then model churn is not a threat. It is just weather. The models become interchangeable suppliers, and you become the thing that does not change.
That reframes the practical decision. Picking an AI video model is not a one-time procurement choice you make and forget. It is a position you hold and revisit, with a backup ready and an exit planned from day one. The teams that will look smart in two years are not the ones that guessed which model would win. Nobody can guess that. They are the ones who built so that the answer did not have to matter.
Council summary
This post argues that picking an AI video model is a strategic bet, not a one-time procurement choice, because the leader can be switched off without warning. OpenAI's Sora 2 is the proof: launched 30 September 2025, the app closes 26 April 2026 and the API on 24 September 2026, with account data deleted after. The council verified the Sora timeline, Runway's 315 million dollar raise at a 5.3 billion dollar valuation, Veo 3.1's January 2026 4K update, the Seedance 2.0 rollout that skipped the US, and the AI video market sizing near 717 million dollars in 2025. We corrected the Sora download milestone to under five days, refined Runway's revenue figures to a verifiable range, and updated Kling's run rate to roughly 500 million dollars by mid-2026. The takeaway for a content team: run a primary model and a credible backup, keep your process model-agnostic, and make editorial judgment, not one tool's prompt tricks, the thing that does not change.
Comments