OpenAI Killed Sora: What the Shutdown Reveals About AI Video Economics — OpenAI Chiude Sora

OpenAI Killed Sora: What the Shutdown Reveals About AI Video Economics

On March 24, 2026, OpenAI announced with 30 days’ notice that it was discontinuing the Sora public API. The reason given was blunt: the economics of AI video generation at scale were “economically irreconcilable” with the prices users would actually pay. Generating one minute of Sora-quality video required compute that cost OpenAI multiples of what API customers were paying. The public API is shutting down. The product lives on inside ChatGPT for paying subscribers, but the developer ecosystem built around it has been told to find alternatives.

This is the most significant AI product discontinuation since GPT-3’s retirement, and it reveals something important about where the AI industry actually is — versus where the hype suggests it should be.

The Economics of AI Video Are Fundamentally Different from Text

Text generation with frontier models costs roughly $2–$25 per million tokens depending on the model. Image generation is more expensive but still manageable at scale. Video generation is categorically different. A one-minute video at Sora’s quality requires rendering thousands of frames with coherent physics, lighting, motion, and scene continuity. The compute cost per generated minute is orders of magnitude higher than text or images — and that cost scales with duration, resolution, and quality in ways that compound quickly.

The fundamental problem is that users expect video generation to cost approximately what stock footage costs — meaning almost nothing, or at most a few dollars per clip. The actual compute cost to generate that footage is far higher. ByteDance paused its video AI rollout earlier in 2026 for related reasons: not just copyright concerns, but economic ones. AI video generation may require a different hardware paradigm — specialized inference chips with much lower cost curves — before the economics become viable for public APIs.

What Sora’s Shutdown Means for Developers

The developer ecosystem that built on Sora’s API has 30 days to migrate. The alternatives are limited: Runway, Kling, and Luma Labs all have video generation APIs, but none match Sora’s quality for complex prompts. Google’s Veo 3 (available via Gemini Advanced) is the closest competitor at comparable quality levels, but it is not available as a standalone API for developers. This creates a gap in the market: high-quality AI video generation at API scale is effectively unavailable as of late April 2026.

For product teams that integrated Sora into their workflows — video content pipelines, creative tools, marketing automation — this is a significant operational disruption. The lesson is one that keeps recurring in the AI product space: building on top of AI APIs requires treating those APIs as infrastructure that can change, deprecate, or shut down, and maintaining fallback options accordingly.

The Broader Pattern: AI Products That Survive vs. Those That Don’t

Sora’s shutdown fits a pattern that is becoming visible across the AI product landscape. Products that survive are the ones where the cost of generation is low relative to the value users receive — text, code, analysis, document processing. Products that struggle are the ones where generation costs are high, where quality thresholds are extremely demanding (users reject anything that looks slightly wrong), and where the competitive moat against alternatives is thin. Video sits at the intersection of all three problems.

The companies investing in AI video hardware — specialized inference chips with dramatically lower cost curves — are betting that this math changes. Until it does, the Sora shutdown is a data point about the real economics of compute-heavy AI generation, not a sign that AI video is a fundamentally bad idea.

---

OpenAI Chiude Sora: Cosa Rivela lo Shutdown sull’Economia del Video AI

Il 24 marzo 2026, OpenAI ha annunciato con 30 giorni di preavviso che stava interrompendo la Sora public API. La ragione fornita era diretta: l’economia della generazione video AI su scala era “economicamente inconciliabile” con i prezzi che gli utenti avrebbero effettivamente pagato. Generare un minuto di video alla qualità di Sora richiedeva un costo computazionale multiplo rispetto a quanto i clienti API stavano pagando.

L’Economia del Video AI È Fondamentalmente Diversa dal Testo

La generazione di testo con modelli di frontiera costa circa $2–$25 per milione di token. La generazione video è categoricamente diversa. Un minuto di video richiede il rendering di migliaia di frame con fisica coerente, illuminazione, movimento e continuità della scena. Il costo computazionale per minuto generato è ordini di grandezza superiore al testo o alle immagini. Gli utenti si aspettano che la generazione video costi circa quanto i filmati stock — quasi nulla. Il costo computazionale effettivo per generare quel filmato è molto più alto.

Il Pattern Più Ampio: Prodotti AI che Sopravvivono vs Quelli che Non lo Fanno

Lo shutdown di Sora si inserisce in un pattern che sta diventando visibile nel panorama dei prodotti AI. I prodotti che sopravvivono sono quelli dove il costo di generazione è basso rispetto al valore che gli utenti ricevono — testo, codice, analisi, elaborazione documenti. I prodotti che faticano sono quelli dove i costi di generazione sono alti e le soglie di qualità sono estremamente esigenti. Il video si trova all’intersezione di tutti e tre i problemi.