OpenAI (Sora, Sora 2)

On February 15, 2024, OpenAI published a blog post and a small set of demonstration videos for a new model called Sora. The videos were stunning. Photorealistic, sixty seconds long, with coherent motion, plausible physics, and complex scenes that no previous model had come close to producing. There were tracking shots through Tokyo streets that looked like they had been filmed by a Steadicam operator. There was a woman walking in front of a neon-lit storefront with reflections in puddles. There was a paper airplane gliding through a forest. The internet exploded. Within hours, people who had been confidently saying that AI video was years away from being usable were having to rewrite their predictions.

Sora was led by a team that included two researchers named Tim Brooks and William Peebles. Peebles, as we discussed in the transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. chapter, was the same Berkeley graduate student whose Diffusion Transformer paper had been rejected from CVPR 2023 for 'lack of novelty.' He had joined OpenAI shortly after publishing the paper, and within fifteen months he was leading the team that built the world's most discussed video model on top of his own architecture. The technical report OpenAI released alongside Sora was titled 'Video generation models as world simulators,' and it described Sora as 'a diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. that operates on spacetime patches of video and image latent codes', which is to say, exactly the architecture from the Peebles paper, extended to handle video by treating time as a third dimension to be patched and attended over.

Sora's first version was technically a research preview. OpenAI did not release it publicly for nearly a year. There was a long period of restricted access for safety testing, red-teaming, content moderation development, and the various other concerns OpenAI takes seriously. The first wide public release came in December 2024 for ChatGPT Plus and Pro subscribers in the US and Canada. The model people actually got to use was meaningfully smaller than the demo version that had been previewed in February, apparently for cost and latencyLatencyThe delay between asking for a generation and getting the first result. Distinct from throughput. reasons, and the quality, while still impressive, did not match the cherry-picked demo clips. This is a recurring theme in generative video: the demos are always better than the production model because the demos are selected, and the production model has to balance quality against the cost of serving millions of users.

Sora 2 launched on September 30, 2025). This was a major upgrade. Audio was the headline feature, Sora 2 generates synchronized audio along with the video, in the same diffusion process, including dialogue with proper lip sync, sound effects timed to actions, and ambient noise that matches the scene. The visual quality also improved significantly. The model generates clips from 10 to 25 seconds long at resolutions up to 1792x1024 in the Pro tier.

Architecturally, Sora 2 is described as a MultimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. Diffusion TransformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. that processes text, images, and audio inputs in parallel streams, with separate transformer paths for each modality joined by attention layers. This is essentially the Multimodal Diffusion Transformer (MM-Diffusion Transformer (DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models.)) pattern from the image side, extended to include audio as a third modality. The training data presumably includes massive volumes of video with paired audio tracks (YouTube, presumably, and similar sources). The fact that OpenAI got this working at scale is technically impressive, joint audio-video generation requires that the model learn the implicit physics of how things sound, which is a much harder problem than learning how they look.

Sora 2 is offered through three channels: a standalone iOS app called the Sora app (which also has social features, you can share your generations and remix others'), the web interface at sora.com, and the OpenAI Application Programming Interface (APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.). Pricing tiers are: ChatGPT Plus at $20/month for limited access, ChatGPT Pro at $200/month for full Sora 2 Pro access with the longest durations and highest resolutions, and per-generation API access for developers.

Here is a piece of news that surprised the field. On March 24, 2026, OpenAI announced that the Sora app and APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. would be shutting down. The app is scheduled to shut down on April 26, 2026, and the API on September 24, 2026. The official messaging has been somewhat opaque, but the implication is that OpenAI is restructuring how it offers video generation, possibly folding it into a broader multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. product within ChatGPT, possibly stepping back from competing in standalone video generation entirely. By mid-2026 the Sora app has gone dark, with the API set to follow in September.

This is a remarkable turn for what was, eighteen months ago, the most-hyped product in generative AI. The most likely interpretation is that OpenAI looked at the unit economicsUnit economicsThe per-unit costs and revenue of a product, for example the cost to generate one image against the price charged for it. of running a video generation service against well-funded competitors (Google's Veo, Kuaishou's Kling, MiniMax's Hailuo) and decided the strategic position was not strong enough. Generating video is expensive, the competition is intense, and OpenAI's strategic priority appears to be its language model business. Whatever the reason, by the time this document is being read, Sora may already be unavailable, and the Sora team's ideas will have to live on through the architectures they pioneered rather than through the product itself.

Lesson for understanding the field: even the labs that pioneered an entire architectural approach can be unable to sustain a product business in it. Video generation is brutal economically. The compute costs are enormous, the user expectations are high, and the competition is fierce. Expect more shakeouts in the next year as labs decide whether they can afford to keep running these services.

Sora came out of a small research group at OpenAI led by Tim Brooks and Bill Peebles, both freshly minted PhDs who had done their doctoral work at Berkeley. The important detail is Peebles. Before OpenAI, he and Saining Xie wrote the 2022 paper that introduced the Diffusion TransformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks., or DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models., which replaced the convolutional U-NetU-NetA convolutional neural network architecture with a characteristic U shape (compress, then expand) used as the backbone of early diffusion models. Largely replaced by Diffusion Transformers in 2023-2024. at the heart of a diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. with a plain transformer operating over patches. That paper is the architectural seed Sora grew from, and, as it turned out, the seed almost every serious video model grew from. Sora was not a lucky product built on borrowed research. It was the same people shipping the thing their own paper predicted.

The name is a small tell about the ambition. Sora is the Japanese word for sky, chosen to evoke open, limitless creative potential. OpenAI revealed it on February 15, 2024 with no waitlist you could join and no product you could touch, just a technical report and a reel of clips that were dramatically more coherent and longer than anything the public had seen. It functioned less like a launch and more like a flag planted: this is where video is going, and we got here first.

The Sora technical report was not titled after a product. It was titled Video generation models as world simulators, and that framing is the whole thesis. The mechanism is worth understanding because the entire field copied it. Sora chops a video into spacetime patches, small blocks that span a little area of the frame across a little slice of time, and treats those patches the way a language model treats tokens. A transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks. is then trained to denoise them. Because patches are a uniform currency, the same model can absorb clips of different resolutions, aspect ratios, and durations, and OpenAI's bet was that if you simply scaled this recipe with enough data and compute, an implicit model of how the physical world behaves would emerge as a side effect of learning to predict pixels.

Part of that bet paid off and part of it was a fundraising narrative wearing a lab coat. What genuinely emerged with scale was striking: a rough sense of object permanence, of three-dimensional consistency as the camera moved, of one action leaving a plausible consequence in the next moment. What did not reliably emerge was physics. Objects still morphed, limbs multiplied, liquids and rigid bodies violated conservation, and cause and effect broke under any real stress. Sora was a spectacular renderer of how the world tends to look and a shaky predictor of how it actually works, and the gap between those two things is exactly where the honest reading of the world-simulator claim lives.

At the two moments Sora mattered, it was genuinely ahead. In February 2024 nothing else came close on coherence and clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. length, and the reveal reset everyone's sense of what was possible. In September 2025, Sora 2 landed a real technical jump: natively generated, synchronized audio rather than dubbed sound, noticeably better physical plausibility, and Cameos, a consent-based feature that let you drop a verified likeness of yourself or a friend into a scene. It shipped alongside a standalone, invite-driven social app that briefly became the most talked-about thing in consumer AI.

The weaknesses are the reason the story ends the way it does. Sora was never the cheapest or the most controllable, and by the time competitors caught the recipe it was rarely the best on independent leaderboards, trailing Google's Veo line and the strongest Chinese models on quality and prompt adherencePrompt adherenceHow faithfully a model follows the specifics of your prompt, meaning the objects, counts, colors, spatial relationships, and text you asked for. High adherence means you get what you described, not just something in the right neighborhood. through late 2025 and 2026. It carried the highest compute burden in the category, it drew sharp objections over mimicking living artists and studio styles, and the social app that was supposed to be its moatMoatA durable competitive advantage that protects a business from being copied or undercut, such as workflow depth, proprietary data, or brand and community. never converted novelty into retention. People came to see the trick, made a few clips of themselves, and left.

The shutdown is best understood as an economics decision, not a technology failure. Running a free-to-cheap consumer video app on frontier models is brutally expensive, with reported compute on the order of a million dollars a day, and the usage curve went the wrong way: an early peak around a million users that then drifted below half that as the novelty wore off. Set a burn like that against a user base that is shrinking, inside a company whose actual mission is artificial general intelligence, and the internal argument writes itself. OpenAI chose to pull the researchers and the compute back toward world models as a research program rather than keep subsidizing a social feed of talking cameos.

So the app went dark and the APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. was put on a clock, but the ideas did not die, they relocated. The world-simulator thesis is now being pursued as serious research inside the big labs, including by the very people who started it, and the DiTDiT (Diffusion Transformer)A diffusion model architecture that uses a Transformer instead of a U-Net. Introduced by Peebles and Xie in 2022. Scales much better than U-Nets and is the basis for almost all current frontier models. plus patches architecture is the water every video model swims in. Sora's other legacy is cultural. It was the moment Hollywood stopped treating generative video as a toy: a major studio expansion was reportedly paused after leadership saw what was coming, artists and studios pushed back hard on style imitation, and the fragility of provenanceProvenanceA verifiable record of how a piece of media was made and by whom, increasingly required so that synthetic content can be traced. was exposed when third-party tools stripped Sora's visible watermark within about a week of release. Sora is the rare product that failed as a business and still moved the entire field, set the technical template, and forced the culture to take the medium seriously.