The death of Sora signals a reset — AI video bowed out, LLM coding agents move to centre stage

App pulled, costs blamed, and a high‑profile experiment ends

It has been reported that OpenAI quietly pulled the Sora app and suspended its developer API, ChatGPT video features and even a reported $1 billion Disney joint‑venture plan — a swift end for an AI video model that briefly rivalled mainstream ChatGPT attention. At launch Sora dazzled with photorealistic video and prompted breathless predictions that AI would upend Hollywood. What killed it? Not lack of novelty, but economics, regulation and copyright risks converging into an unsustainable business.

Sora’s metrics, reportedly, tell the story: downloads spiked to 3.3 million soon after release and later fell to about 1.1 million, in‑app revenue was roughly $2.1 million while OpenAI’s daily spend on video generation was said to be about $15 million — a run‑rate that would exceed $5 billion annually. OpenAI executive Bill Peebles reportedly acknowledged the model’s economics were untenable. Compounding the cost problem were deepfake risks and tightening rules: China has issued AI video labeling and authorization measures, and ByteDance (字节跳动) paused global rollout of its Seedance 2.0 amid unresolved training‑data copyright issues.

A strategic pivot: from flashy video to code, agents and enterprise

The immediate consequence is strategic. OpenAI has reportedly redirected resources back to language models and coding agents — Codex and a forthcoming model called Spud are front and centre — while CEO Sam Altman said the move frees capacity for “next‑generation” AI work. That shift aligns with a broader industry split: Google’s multimodal push (Veo in Gemini) and ByteDance’s Seedance still chase consumer video, but many leading labs are doubling down on LLMs-as‑cognitive‑substrates that act through tools, code execution and memory. Anthropic, long focused on language‑first approaches, has been cited as an example of a lab that turned language models into robust agentic systems rather than pursuing brute‑force multimodal video.

This repositioning is partly philosophical. Critics such as Yann LeCun argued that pure text prediction cannot reach embodied intelligence; proponents of the language‑agent route answer that LLMs gain grounding when embedded in perception‑action loops, compiler‑verified code execution and iterative feedback — areas where models like Claude Code have scored strongly on software engineering benchmarks. The market test of Sora, analysts say, shows that spectacular media demos are not the same as scalable, safe, value‑creating AI businesses.

Broader implications: compute, policy and ethics

Sora’s fall matters beyond product strategy. It recalibrates demand forecasts for compute and inference hardware — an outcome with geopolitical resonance as governments scrutinize exports of advanced GPUs and the design of compute supply chains. It also revives ethical debates: should resource‑hungry, deepfake‑capable models be open‑sourced for community benefit (a suggestion made publicly by Hugging Face’s founder), or should they be curtailed because social harm outweighs experimental value? For Western readers unfamiliar with China’s ecosystem: domestic rivals and regulations (e.g., ByteDance (字节跳动)’s Seedance 2.0 and national labeling rules) will shape where the video lane survives and where it does not.

Sora’s demise is, in short, a market verdict. The spectacle of AI video has been checked by money, law and data rights. What comes next is a contest over which form of LLM‑based intelligence — multimodal media engines, tool‑using coding agents, or world‑model initiatives — best converts technical prowess into scalable, governable and economically viable products.