news

Why OpenAI walked away from Sora, and why video specialists keep winning

Sora was a competitive model. It still lost the fight for compute inside a company whose main business is somewhere else.

By Daniel Ochoa · 3 October 2026

Sora 2 was not a weak model. In our benchmark it scores 4.78 out of 5, sixth of the 12 models we have tested and only 0.13 points behind the leader, with full marks on faces, multi-shot continuity and lip-sync. OpenAI shut it down anyway. The API was retired on September 24, 2026, after a wind-down announced in March and laid out in OpenAI's discontinuation notice. Our migration guide covers what to switch to.

A side project competing for GPUs

Reported figures put Sora's compute bill at roughly $1 million a day, against subscription revenue estimated at under $2 million a month, while usage slid from near one million monthly users to under half that. For a generalist lab, every GPU-second spent on video is a GPU-second not spent on the products that pay the bills, and a GPU-second spent on text usually brings in more revenue than one spent on video. When OpenAI needed compute elsewhere, video was the obvious thing to cut.

That is the structural case for specialists. When video is the whole business, it cannot be deprioritized to free up compute for something else, and the whole company is tuned around it: serving costs, distillation, controls, and the editors and pipelines their customers work in. Google is the clear exception among the generalists, putting Gemini Omni, currently first on our board, straight into its own products, including a free tier in Google Vids.

Being a specialist is not a moat either

If base models converge, and our data says they are converging at the top, then a specialist whose only edge is a slightly better model is exposed too. Open weights shorten any lead further, although openness in video is inconsistent:

• Alibaba released Wan 2.1 and Wan 2.2 under the Apache 2.0 license, then kept every later Wan generation behind its API.
• MiniMax published open weights for H3 on August 3, 2026, but its license bars self-hosting in the United States, the European Union, the United Kingdom and South Korea without separate permission. MiniMax H3 scores 4.81 in our benchmark, fifth overall.

Even so, companies can now run a competitive open-weight video model on their own hardware, subject to those terms. Owning the best model is a weaker position than it was a year ago.

Where the durable value is moving

Look at who sells the models and the pattern is clear. On our price radar, fal, Replicate, Runway and Higgsfield each resell several of the same frontier models. Runway, which builds its own Gen-4.5, also sells Veo, Seedance, MiniMax H3 and Grok Video, a shift we covered in Runway's multi-model pivot. Routing between models on its own is easy to copy, so the advantage has to come from what sits around the model: control over identity and multi-shot sequences, review and approval, checks that output meets the brief, provenance, and the production data a platform gathers from its customers' approvals.

Leaders are betting on different growth paths: interactive and real-time generation, native audio and video in one model, or world models for robotics and simulation. Each is credible. None of them is "have the best clip generator", which is the lesson Sora's exit makes hard to ignore.

Frequently asked

Why did OpenAI shut down Sora?

OpenAI announced the wind-down in March 2026 and retired the API on September 24. Reported figures put Sora's compute cost at about $1 million a day against far smaller revenue, at a time when the company was putting its compute into other products.

Was Sora 2 a bad model?

No. It scores 4.78 out of 5 in our benchmark, sixth of 12 and within 0.13 points of the leader, with perfect marks on faces, multi-shot continuity and lip-sync.