OpenAI killed Sora. The math explains why.

Six months, start to finish

OpenAI previewed Sora in February 2024. The actual consumer launch didn't happen until December 2024, bundled into ChatGPT Plus. Sora 2 followed in September 2025 with a standalone iOS app that hit 1 million downloads in under five days. Faster than ChatGPT did.

By October, Bill Peebles - the head of Sora - was already saying the quiet part out loud on X. He called the economics "completely unsustainable." November was peak: 3.3 million downloads in a single month. Then the curve bent. December dropped 32%. January dropped another 45%. By February 2026, downloads were at 1.13 million, down two-thirds from peak.

OpenAI revoked free-tier access on January 10, 2026. That accelerated the decline but probably wasn't the cause. The product just didn't retain people. WSJ described the output as "more AI slop than AI magic," which is brutal but consistent with what we saw. Most Sora videos looked impressive for about three seconds before something weird happened to someone's hands or a dog's legs.

On March 24, 2026, OpenAI announced the shutdown. The app closes April 26. The API stays on life support until September 24, then goes dark.

Why every clip cost OpenAI $1.30

Cantor Fitzgerald analyst Deepak Mathivanan broke this down and the math is rough. Each video generation took about 40 minutes of total GPU time - four GPUs running in parallel for 8-10 minutes each. At roughly $2/hour per GPU, that's about $1.30 per 10-second clip.

Now multiply that. Mathivanan estimated around 4.5 million app users, with maybe 25% generating roughly 10 videos per day. That comes out to 11.3 million clips per day. At $1.30 each, you get the $15 million daily inference cost figure that Forbes reported.

That $15M/day number is the gross compute cost at peak usage. The WSJ reported a different figure - about $1M/day in net operating losses - which likely reflects costs after revenue, and probably came from a period when usage had already dropped. Both numbers are directionally right, just measuring different things.

Either way, the math didn't work. $2.1 million in total lifetime revenue against even the lower $1M/day loss figure means Sora burned through its entire revenue in about two days of operations. Against the peak $15M/day inference cost, the total revenue covered roughly 8 minutes.

What everyone else charges

Sora was expensive to run and expensive to use. Competitors figured out how to be at least one of those things cheaper.

Pika and Kling run at 7-17 cents per video. Sora cost OpenAI $1.30 to produce each clip. The competitors were outpricing Sora by 7-18x on production costs alone, even before you factor in that their subscription plans started at $7-12 instead of the $20 ChatGPT Plus minimum.

The Disney deal that wasn't

In December 2025, Disney and OpenAI announced a partnership. Disney would invest $1 billion in OpenAI and license characters - Mickey Mouse, Darth Vader, the whole catalog - for use in Sora. Bob Iger personally signed off on it.

No formal agreement was ever signed. No money changed hands.

When OpenAI announced the Sora shutdown on March 24, Disney reportedly learned about it less than an hour before the public did. Variety and The Hollywood Reporter both described Disney as "shocked." The diplomatic press statement that followed was Disney at its most controlled. This was not how a billion-dollar partnership is supposed to unwind.

What developers were paying

The Sora API had two tiers. Standard Sora 2 ran $0.10 per second at 720p. Sora 2 Pro cost $0.30/second at 720p and $0.50/second at 1024p (1792x1024). Both supported 4, 8, and 12 second clips. Pro could do up to 25 seconds.

One detail that frustrated developers: every generation attempt charged full price, whether the output was good, got flagged by the content filter, or just timed out. You paid $1 for a 10-second 720p clip even if the clip was unusable.

Compare that to text LLMs where a failed generation costs fractions of a cent. The risk per API call was fundamentally different. A text model that hallucinates costs you maybe $0.003. A Sora generation that produces garbage costs you a dollar.

Video generation is a different economics game

We track 400+ text LLM prices and the trend is clear: costs drop roughly 50% every six months. GPT-4 launched at $30 per million tokens in March 2023. Today you can get equivalent quality for under $0.15.

Video doesn't follow the same curve. Text generation is fast - tokens stream at 50-200 per second on most models. A $0.003 API call takes a second or two. Video generation needs 40 minutes of GPU time across multiple cards to produce 10 seconds of output. The compute intensity is orders of magnitude higher.

Text LLM inference has also benefited from years of optimization: speculative decoding, quantization, KV-cache reuse, MoE routing. Video diffusion models haven't had the same level of inference engineering attention yet. The $1.30/clip cost that killed Sora will come down, but it probably won't hit $0.01 in the next year the way text models did.

There's also a demand problem. People expect AI text to cost nearly nothing now. They expect AI video to cost nearly nothing too. But the compute gap between the two is something like 10,000x per output. Sora was caught between user expectations shaped by cheap text models and infrastructure costs shaped by physics.

Spud: the pivot nobody asked about

OpenAI is redirecting the Sora team toward what Bill Peebles described as "systems that deeply understand the world by learning to simulate arbitrary environments at high fidelity" - basically, world simulation for robotics. Separately, OpenAI has a new model codenamed "Spud" in development, though that appears to be a general frontier model rather than a Sora-specific successor.

Peebles called the prize "automating the physical economy." The robotics pivot will be API-only, not a consumer app. That's probably the right call given what just happened.

The subtext is interesting. OpenAI spent two years building the best consumer AI video product, discovered the economics were impossible at consumer prices, and is now pivoting to enterprise/industrial use cases where they can charge more per unit. Whether that works depends on whether industrial customers have video generation problems worth $1.30+ per clip. Some of them might.

What this tells you about AI costs in general

Sora is a useful data point for anyone budgeting AI infrastructure. A few things are worth keeping in mind.

Not all AI modalities are priced the same, and they won't converge anytime soon. Text generation at $0.15 per million tokens is a completely different economic product than video generation at $1.30 per clip. Image generation sits somewhere in between - DALL-E 3 is about $0.04 per image, and Midjourney runs roughly the same per generation. Audio (TTS and STT) falls closer to text pricing. Each modality has its own cost curve.

The text LLM price decline we've been tracking is specific to text. It happened because of competition (DeepSeek, Google undercutting), architectural improvements (MoE, distillation), and hardware upgrades (H100 to B200). Video generation hasn't had its GPT-4o-mini moment yet. When it does, the economics will shift. Until then, budget accordingly.

If you were building on the Sora API, you have until September 24 to migrate. Kling at $0.07/second is the cheapest API option for video right now. Google Veo 2 via Vertex is another path, though pricing varies by configuration. Both have different quality profiles than Sora. Test before you commit.