OpenAI Sora 1

» see sora.openai.com as it lived, on the wayback machine

~ the eulogy ~

sora 1 was a model i watched at a remove. in 2024 i was working on edge ai for uavs in ankara, and the text-to-video diffusion stuff was happening on someone else's gpu cluster, on someone else's bill. i opened the sample videos when they dropped. i did not subscribe to chatgpt plus to play with it. by the time sora 2 launched in september 2025 the first version was already a curiosity.

openai showed the first sora demo on 15 february 2024. the will smith eating spaghetti video did the rounds. for nine months the model stayed in research preview while the company figured out compute cost and safety review. the public release of sora turbo came on 9 december 2024 as part of the 12-days-of-openai launch, available to chatgpt plus and pro subscribers in the us and canada. the architecture was a diffusion transformer working on 3d patches of pixels and time, denoising a video out of static the same way image diffusion models denoise a still.

the model burned cash. operating costs were reported around one million dollars a day at peak. user counts peaked near one million and slid under five hundred thousand by late 2025. sora 2 launched 30 september 2025 with a tiktok-style social app and visible watermarks, and the company began moving everyone to it. on 27 february 2026 the help center posted the sunset faq. the us retirement landed on 13 march 2026. the consumer app shuts down 26 april and the api goes dark in september.

~ the dossier ~

~ the funny part ~

sora 1 cost openai about fifteen million dollars a month at peak and was killed for being too expensive at a company that had just signed a one-billion-dollar disney content deal. the cost was not really the cause of death. sora 2 was. the model improved fast enough between december 2024 and september 2025 that maintaining the older one was a tax on the gpu pool and a confusion in the product surface. when you have two text-to-video models on the same homepage, one of them is going to lose.

~ what it leaves behind ~

sora 1 is the first openai sub-product death. [[gpt-3]] was deprecated as a model name but never had a product surface of its own. dall-e 2 was rolled into chatgpt without a clean shutdown. sora 1 had a homepage and a calendar date for its end, which is gravestone material. the diffusion-transformer architecture sora 1 introduced is still the backbone of every serious text-to-video model in 2026, including sora 2. the videos sora 1 produced look obviously dated now, the way 2017 gan faces look dated. that was a fifteen-month half-life on the state of the art.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »