Sora and Veo (Video Generation Models)

Sora (OpenAI) and Veo (Google DeepMind) were, through 2024–2025, the two most prominent frontier text-to-video model families. Both are diffusion transformers with undisclosed parameter counts and training data, and both were positioned by regulators and litigants as the canonical referents for synthetic-media and deepfake policy, including the TAKE IT DOWN Act, state deepfake statutes, and the synthetic-media provisions of the EU AI Act. They are described as the tools most likely to produce convincing synthetic media for political, financial, or intimate-imagery harm cases, which gives their policy relevance a profile larger than their market share. In March 2026 OpenAI announced it would shut down Sora as part of a refocus on business and productivity applications, while Google continued to invest in the Veo line (Source: arstechnica.com; 9to5google.com).

OpenAI Sora

The Sora 1 preview (Feb 15, 2024) introduced 60-second coherent video generation and was widely described as an inflection moment for video AI capability; competitors including Runway and Pika compressed their timelines in response. The general release (Dec 9, 2024) shipped in the ChatGPT Plus ($20) and Pro ($200) tiers, restricted to a 20-second maximum duration and up to 1080p, with C2PA metadata and a visible watermark applied.

Sora 2 (Sep 30, 2025) added native synchronized audio including dialogue and SFX, raised output length to 15 seconds, improved physics plausibility, and introduced a "cameo" feature allowing likeness-based personalization with consent controls. It launched alongside a dedicated Sora app, which was briefly ranked #1 on the iOS App Store.

Sora's safety regime refuses generation of public-figure likenesses without explicit consent mechanisms, refuses NCII-adjacent content categories, embeds C2PA provenance metadata, and applies a visible watermark on all free and Plus outputs. C2PA does not yet have a standalone page and is relevant to the Synthetic Media / Deepfakes coverage gap. Public criticism has focused on the ease with which watermarks can be cropped, the "cameo" feature as a normalization of likeness-AI, and training-data disclosure: OpenAI has not specified whether licensed video or scraped YouTube and commercial footage is in the corpus (Source: AI Copyright Litigation — Analysis).

Discontinuation

On March 24, 2026, OpenAI announced it would shut down Sora, roughly fifteen months after the app's public launch. The announcement came in a social-media post shortly after the Wall Street Journal reported the decision; OpenAI said it would share "timelines for the app and API and details on preserving your work," and thanked users who "created with Sora, shared it, and built community around it." The move followed a reported all-hands meeting at which executives said the company was refocusing on business and productivity applications rather than being "distracted by side quests," as head of applications Fidji Simo reportedly put it. It also came months after Disney invested $1 billion in OpenAI in December 2025 in a deal licensing around 200 Disney characters for use in Sora; how that partnership would continue was not clear at the announcement (Source: arstechnica.com). By late June 2026, Sora was no longer generally available (Source: cnet.com).

Google Veo

Veo 2 (Dec 2024) supports up to 4K and 2-minute generation via the API, with high motion coherence. Veo 3 (May 2025) added native audio generation including synchronized dialogue, the headline advance over both Veo 2 and Sora 1, and is available in the Gemini app, Vertex AI, and Google's Flow video-editing product. Veo 3.1 (2025 updates) brought further quality improvements and extended Veo to YouTube Shorts creator tools through the Flow integration.

Veo's safety regime applies SynthID invisible watermarking on all outputs, Google's proprietary watermark system, which Google describes as more robust than visible-only approaches. It refuses public-figure likenesses, explicit content, and violent content, and integrates with Google's ads and YouTube content policies.

On March 31, 2026, days after OpenAI's Sora shutdown announcement, Google said it remained committed to offering video generation and announced Veo 3.1 Lite, a lower-cost tier (Source: 9to5google.com) that roughly halved API generation costs relative to prior Veo pricing (Source: decrypt.co). The Veo models also form the technical basis of Google's Genie world models (Project Genie, January 2026), which allow a degree of real-time interactivity with generated video content (Source: arstechnica.com).

Technical notes

Both Sora and Veo are diffusion transformers operating in compressed latent space. OpenAI's Sora technical report describes spatiotemporal patches used as tokens, training on videos of variable resolution, duration, and aspect ratio (with no fixed 16:9 256×256 constraint), and a re-captioning pipeline applying dense synthetic captions from a GPT-4V-class model to the training corpus. Parameter counts, training compute, and dataset composition are undisclosed for both families.

Market entry and reception

Sora's Feb 2024 preview is credited with forcing Meta, Google, and several startups to accelerate their video roadmaps, with catalyzing the Hollywood guilds' 2024 bargaining position on AI-generated content, and with triggering the SAG-AFTRA "AI Safeguards" provisions. Veo 3 (May 2025) was described at release as the technical frontier, particularly after native audio added the dimension previously missing from text-to-video output.

The competitive landscape shifted substantially in 2025–2026. ByteDance's Seedance 2.0 drew attention for viral videos of complex, Hollywood-style scenes with multi-shot cuts and camera angles (Source: arstechnica.com). By June 2026, Alibaba Cloud's HappyHorse 1.1 had risen to No. 2 in global video-model rankings as Sora and Seedance fell in the standings (Source: venturebeat.com).

Deepfake and synthetic-media relevance

Sora and Veo recur as the frontier referents across several policy instruments. Under the TAKE IT DOWN Act, the US federal NCII and deepfake law enacted in 2025, deployment of video-generation tools is part of the threat model. Under EU AI Act Article 50, which sets transparency obligations for synthetic media, both providers are designated GPAI with systemic risk. At the state level the relevant statutes include California SB 942 (AI transparency), the Minnesota and Washington deepfake statutes, Texas HB 18, and New York's election-deepfake statutes.

In AI First Amendment debates, generating plausible political speech in a specific likeness is treated as the hard case for speech protection; the public-figure refusals built into Sora and Veo are characterized as a voluntary-restraint response that regulators may not accept as permanent. Training-data opacity remains the open policy question linked to AI copyright litigation. The synthetic-media and deepfake concept page is not yet created (see Missing Coverage: Companies, People, Models, Legislation, Concepts (revised) and the Synthetic Media gap). Both families are the flagship multimodal generation products of OpenAI and Google DeepMind.

Relationships

• instance-of: General-Purpose AI (GPAI)
• related: OpenAI, Google DeepMind, GPT-4 Family (OpenAI), Gemini 3 / Gemini 3 Pro, AI Copyright Litigation — Analysis, EU AI Act (Regulation 2024/1689), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment)