Open AI Video Generation: What It Does, What It Costs, and When It's Worth It

The Sora Story: What OpenAI's Video Model Could Actually Do

When OpenAI launched Sora 2 on September 30, 2025, it promised a "GPT-3.5 moment for video." The hype was immediate and massive. Within days, the app pulled in over a million downloads, reaching four million by Halloween. At its peak, the system generated approximately 11.3 million videos daily. This surge demonstrated the intense market appetite for high-fidelity AI motion, yet it also exposed the severe economic realities of current inference costs.

The technical capabilities were genuinely impressive for a consumer-facing tool. Sora 2 introduced synchronized dialogue, sound effects, and background soundscapes that matched the visual action. It also improved physical accuracy, ensuring objects like basketballs rebounded realistically rather than teleporting. A key feature called "Cameos" allowed users to insert their own verified appearance and voice into scenes. However, the model still struggled with object permanence and long-range coherence, often failing to maintain consistent world states across multiple shots.

• Synchronized audio and dialogue aligned with visual movements
• Improved physical accuracy in object interactions and spatial relationships
• Multi-shot instruction following with persistent world state attempts
• Cameos feature for inserting user-verified identities into scenes

Despite the technical strides, the underlying economics proved catastrophic for the business model. Analysts estimated the internal compute cost at approximately $1.30 per 10-second clip. This figure starkly contrasted with consumer pricing structures that often hovered around $0.40 per video in bulk packs. The disparity was so significant that OpenAI’s head of Sora admitted the economics were "completely unsustainable." Total lifetime in-app revenue reached only about $2.1 million, a fraction of the estimated daily operating costs.

This financial mismatch led to a rapid decline in adoption and eventual shutdown. Downloads fell sharply after the initial hype, with retention rates dropping to single digits. OpenAI officially discontinued the Sora product, with the API removal scheduled for September 24, 2026. The platform served as a critical learning phase for the industry, highlighting that free or cheap consumer AI video is not yet economically viable at scale. Businesses must now treat Sora-specific details as historical context rather than active options.

For enterprise teams, the lesson is clear: AI video remains valuable for specific use cases like training, demos, and internal communications. These applications align better with current cost-per-output math than mass consumer generation. As you evaluate AI tools for your workflow, consider how agent skills can automate surrounding steps like prompt drafting and clip stitching. You can browse and test free agent skills on SkillzBench to streamline these processes. This approach ensures you leverage AI video where it adds genuine value, without getting caught up in the hype of discontinued consumer products.

What It Cost: The Pricing Tiers Nobody Could Make Sustainable

Sora 2 looked like a bargain on paper: 1,000 credits for $20 a month, or 10,000 credits at the $200 Pro tier, with API access starting at $0.10 per second of 720p video. The catch was that almost nobody generated one video per prompt.

Users typically needed 3 to 7 attempts before landing an acceptable clip, and every failed attempt burned credits anyway, according to a breakdown of Sora 2's paid plans. Peak-hour error rates (9am–5pm PST) ran 35–42%, versus just 3–8% off-peak — meaning the busiest, most expensive hours were also the least reliable. A "cheap" $20 plan could quietly cost several times its sticker price in wasted generations.

The deeper problem was the unit economics underneath the pricing page. Analysts estimated an internal compute cost of roughly $1.30 per 10-second clip — about 40 minutes of GPU time at just under $2/hour — against consumer pricing of roughly $0.40 per video in bulk packs, per reporting on Sora's cost structure. OpenAI's own head of Sora, Bill Peebles, called the economics "completely unsustainable."

The tiers, in hindsight, were never built to balance:

• Free invite tier — viral growth, zero revenue, maximum compute burn
• ChatGPT Plus at $20/month — 1,000 credits, 720p, watermarked; roughly six 10-second clips per month
• ChatGPT Pro at $200/month — 10,000 credits, 1080p, up to 25 seconds, priority queue; heavy users saw per-video costs fall to about $0.80
• API access — $0.10/second at 720p standard, $0.30–$0.50/second for Pro tiers
• Extra generations — about $4 for 10 additional videos, with a 24-hour download window

The verdict arrived in the numbers. Sora's total lifetime in-app revenue reached only about $2.1 million, against operating costs estimated at roughly $1 million per day — a gap that enterprise analysts noted could never close. Sam Altman put it bluntly: there is "no ad model" that supports people making memes for three friends.

For teams planning around AI video today, the lesson is to budget for iteration, not list price — and to treat any per-second rate as a floor, not a total. Agent skills can automate the surrounding workflow, from prompt drafting to clip review; SkillzBench is building a free directory and MCP connector to help teams find and test the ones that actually deliver.

When AI Video Is Actually Worth It: Matching Use Cases to Economics

Sam Altman put the consumer problem bluntly: people were generating funny memes to send to three friends, and "there is no ad model that can support the cost of that kind of a world." That admission, reported during Sora's run, explains why a product with 4 million downloads in its first month couldn't survive — and why the smart question for businesses isn't "should we use AI video?" but "for what, exactly?"

The consumer math never worked. Each 10-second clip cost OpenAI roughly $1.30 in compute against consumer pricing near $0.40, and lifetime in-app revenue topped out around $2.1 million. When the app shut down and the Sora 2 API removal was scheduled, the message was clear: entertainment-scale generation is a subsidy, not a business.

Enterprise use is different, because the economics and quality bar both change. Constellation Research analyst Liz Miller reviewed Sora alongside 22 enterprise video solutions and found the output "entertaining but often unusable for final production" — yet genuinely valuable for storyboarding and concepting, where a rough draft beats a blank timeline. That's the pattern worth internalizing:

• Training and internal comms, where speed and clarity matter more than cinematic polish
• Product demos and pitch decks, lowering the cost of showing an idea before building it
• Storyboarding and concepting, treating AI output as draft material, never final deliverables
• Previsualization for creative teams, iterating cheaply before committing production budget

Two practical cautions apply. First, budget for iteration: users typically needed 3–7 attempts per acceptable clip, with every failed attempt consuming credits, though structured prompt templates lifted first-try success to 87% and short 5-second clips succeeded 99%+ of the time. Second, govern the output. NewsGuard found Sora 2 produced videos with provably false claims in 16 of 20 prompts — an 80% rate — and brand logos weren't covered by guardrails, so any business use needs human review and brand-safety checks.

For production-adjacent work, enterprise analysts point toward commercially safer tools such as Adobe Firefly, Google Veo 3.1, ElevenLabs Video, and Leonardo.ai, which offer clearer licensing and watermarking (per Constellation's enterprise review). Costs are also falling fast — analysts project video inference prices dropping roughly fivefold by next year and threefold more by 2027 — so revisit the math quarterly.

The surrounding workflow matters as much as the model: prompt drafting, clip stitching, review routing, and publishing can all be automated. Agent skills can handle those steps, and SkillzBench is building a free directory and benchmarking platform so you can see which skills actually help before wiring them into your agents — worth watching as AI video economics improve.

Your Practical Playbook: Budgeting, Prompts, and What to Watch Next

Sora's collapse taught a hard lesson: the sticker price of an AI video tool is never the real price. The teams that got value were the ones who planned for iteration, kept clips short, and treated every output as a draft.

Start by budgeting for attempts, not subscriptions. Even when Sora was live, users typically needed 3–7 attempts per acceptable video, and every failed attempt consumed credits, according to pricing analyses from the period. Multiply your expected output volume by that retry factor before committing to any tier.

Then, shrink your first experiments. Independent testing data found that 5-second clips succeeded over 99% of the time, while complex prompts failed more often than they worked. A structured prompt template pushed first-try success to 87%. Your evaluation playbook should look like this:

• Budget 3–7x your list price for iteration on every video goal.
• Prototype with 5-second clips before attempting longer sequences.
• Standardize a structured prompt template rather than free-form descriptions.
• Treat all output as draft-quality — best for storyboarding and concepting, not final production, as enterprise analysts concluded.

Governance matters more than most teams expect. NewsGuard testing found Sora 2 produced videos with provably false claims 80% of the time when prompted, and brand logos fell outside existing guardrails. Any business deployment needs brand-safety review, watermarking policies, and human sign-off before anything ships externally.

Two signals should keep AI video on your roadmap despite Sora's shutdown. Sam Altman has hinted that video generation could return as a ChatGPT feature rather than a standalone app, and analysts project video inference costs dropping roughly fivefold within a year. The unit economics that killed Sora are improving fast — the underlying capability isn't going away.

Meanwhile, the surrounding workflow doesn't have to be manual. Agent skills can automate prompt drafting, clip stitching, and publishing, and you can browse and test free ones — along with the free MCP connector that loads them into your agent — on SkillzBench's directory as it prepares for launch. Revisit video generation when costs fall; automate the pipeline around it now.