Three AI Video Models, Three Buyers: Omni vs Sora vs Veo 3

Earlier this week, on May 19, 2026, Google announced Gemini Omni Flash at I/O 2026 — the first member of its any-input multimodal Omni family, accepting text, image, audio, and video as input and generating up to 10 seconds of video output, freely available inside YouTube Shorts Remix. That launch closed a 23-day consumer vacuum left when OpenAI's Sora app died on April 26, 2026. Now three production video-generation platforms are operating in parallel, each aimed at a different buyer.

The stakes for choosing wrong are real. Per-second pricing across the three platforms spans a verified 17× range — from Veo 3.1 Lite at $0.03/sec to Sora 2 Pro at $0.50/sec for 1080p output. Sora 2's API carries a hard shutdown date of September 24, 2026, meaning any team building on it today has a four-month migration window. And Google's two video products — Omni Flash and Veo 3.1 — deliberately co-exist on different surfaces rather than replacing each other.

This guide covers the full 3-way matrix: what Gemini Omni Flash actually shipped, where Sora 2's API still lives and when it dies, why Veo 3.1 is the risk-adjusted enterprise pick, a side-by-side comparison matrix across nine dimensions, the complete $0.03-to-$0.50 pricing ladder, the watermarking landscape (SynthID vs C2PA vs Sora's visible mark), and a routing guide that maps each product to its natural buyer. For the broader post-Sora video-generation landscape including Runway and Kling, see our post-Sora AI video generators guide.

Most launch-week coverage of AI video models asks "which is best?" That is the wrong question. The better question — the one that drives an actual procurement or engineering decision — is: which model were you ever going to ship with? The answer depends almost entirely on your buyer profile, not on VBench scores or vendor self-reports.

No independent third-party benchmark has scored Gemini Omni Flash, Sora 2 Pro, and Veo 3.1 in a single matched run as of May 22, 2026. Each vendor self-reports preferred metrics: OpenAI called Sora 2 “the GPT-3.5 moment for video,” and Demis Hassabis called Veo 3 “state-of-the-art.” Both are vendor claims and should be treated as such. What is objectively verifiable is the pricing, the deployment surface, the modality support, and the API lifecycle status — and those four dimensions are more useful for decision-making than any benchmark.

Google also took care to confirm that Omni Flash and Veo 3.1 deliberately co-exist: the I/O 2026 announcements continued to list Veo 3.1 alongside Omni Flash, with Veo handling video-first generation on Vertex AI and Omni handling any-input multimodal generation in the consumer app. Do not interpret Omni Flash as a Veo replacement — or as “Sora's replacement.” It is a new multimodal family launched alongside an existing video-specialist product. For context on where the broader video market is heading after the Sora shutdown, see our AI video market after the Sora shutdown analysis.

Gemini Omni Flash launched at Google I/O 2026 on May 19, 2026, as the first shipping member of Google's “any-input multimodal” Omni family. The product accepts text, image, audio, and video as inputs and generates up to 10 seconds of video output. As of May 22, Omni Flash is live in the Gemini app, Google Flow, and the YouTube Shorts Remix surface — the developer API was described as “coming in the next few weeks” and is not yet live.

Koray Kavukcuoglu, CTO of Google DeepMind, described the launch as: “Today we're introducing Gemini Omni — a new generation of multimodal models that creates anything from any input, starting with video.” Nicole Brichtova, Product Management Director at Google DeepMind, clarified to TechCrunch that the 10-second output cap is not a model limitation but “a decision based both on a desire to get it into more hands.”

The output resolution is reported at 720p / 24fps by third-party coverage — Google did not publish a resolution specification at launch. Every Omni output carries an imperceptible SynthID watermark; it cannot be disabled via the API. Subscription access runs through Google's AI plan tiers: AI Plus ($7.99/mo), AI Pro ($19.99/mo), and AI Ultra ($200/mo, cut from $250 at I/O 2026). Flow credits meter Omni access for video generation — 200 credits/mo on AI Plus, 1,000 on AI Pro, 10,000+ on AI Ultra.

The strategic insight is the free YouTube Shorts surface. Sora required a $20/month ChatGPT Plus subscription. Omni Flash is free for Shorts creators 18 and older via the YouTube Shorts Remix and YouTube Create interfaces. That pricing delta — $20/mo versus free — is the structural explanation for why Sora's consumer app could not survive once Omni arrived. For more on the AI video creation tools available inside YouTube Shorts, see our guide to Veo 3 Fast inside YouTube Shorts.

OpenAI launched Sora 2 on September 30, 2025, with synchronized audio, dialogue, and sound effects generated in a single pass. Bill Peebles, Head of Sora at OpenAI, described it on the Sequoia “Training Data” podcast in October 2025: “We've hit this kind of like GPT-3.5 moment for video. Let's make sure the world is kind of aware of what's possible now.” The Sora Team's own launch post framed Sora 2's edge in terms of physics simulation: “In Sora 2, if a basketball player misses a shot, it will rebound off the backboard” — contrasting with prior models that would spontaneously teleport the ball to the hoop.

Six months later, OpenAI discontinued the Sora consumer app and sora.com on April 26, 2026. The banner on the Sora 2 launch page now reads: “As of April 26, 2026, the Sora product is no longer available.” The economics behind this decision — the cost structure that made the consumer app unsustainable — are covered in depth in our analysis of the $1M/day economics that made Sora's app unsustainable.

What survives is the API. The Decoder confirmed the two-stage shutdown: app first (April 26), API second (September 24, 2026). The model IDs sora-2 and sora-2-pro remain callable through September 24. API duration parameters are discrete — Sora 2 accepts 4, 8, and 12 seconds; Sora 2 Pro adds 15 and 25-second clips. Any engineering team running programmatic video generation against the Sora API has approximately four months to migrate. The full picture of why OpenAI shut down the consumer app and what it means for the product-market-fit lessons of AI video is covered in our post on why OpenAI killed the Sora consumer app after six months.

Veo 3 originally shipped at Google I/O 2025 (May 2025) with native audio generation. Veo 3.1 launched on October 15-16, 2025 as the upgrade — the current flagship SKU. Demis Hassabis announced it on X: “Veo 3 is the state-of-the-art in video models. Veo 3.1 is our new big upgrade with enhanced realism, richer audio, scene extension, better narrative control, more precise editing capabilities & much more.” The Veo 3.1 Lite tier launched March 31, 2026, completing the three-tier stack: Lite, Fast, and Quality.

Veo 3.1 generates up to 8-second clips at 1080p (the Quality tier supports 4K), with native audio — dialogue, sound effects, and ambient music — in a single generation. Scene extension allows continuing an existing video clip, a capability neither Omni Flash nor the base Sora 2 model supports. The deployment surface spans Vertex AI, the Gemini API (AI Studio), Google Flow, the Gemini app, Google Vids, and Google Photos. Enterprise teams on Vertex AI get full programmatic access with documented pricing and no announced sunset.

The risk-adjusted case for Veo 3.1 is straightforward. Of the three platforms in this comparison, only Veo 3.1 has been in stable production for seven months with no shutdown rumor, no API sunset clock, and no consumer-app drama. Sora 2's app was killed after six months. Omni Flash's developer API is not yet live. For enterprise brand video workflows that need programmatic access, reliability guarantees, and Google Workspace integration, Veo 3.1 is the lowest-variance option available in May 2026. For context on how Veo 3 Fast operates inside YouTube Shorts — the consumer-creator surface — see our YouTube Shorts AI creation guide. For Veo-powered Performance Max video ads, see our guide to Veo-powered Performance Max video.

Demand signals from Veo's May 2025 launch suggest the appetite is real. Hassabis posted in May 2025 that “millions of videos have been generated in the past few days alone” after the original Veo 3 launch. Veo 3.1 does not have an equivalent published figure from Google — but the product has been available to Vertex AI customers for seven months without a major incident.

The table below is the central deliverable of this post. No launch-week comparison as of May 22, 2026 has assembled all three of Omni Flash's launch-day surface, Sora 2's pricing and sunset timeline, and Veo 3.1's current tier matrix in a single nine-dimension comparison. The audio-input row (Omni only) and the scene extension row (Veo only) are particularly differentiated — most comparisons omit both. Pricing figures for Sora 2 are from third-party API aggregators (Apiyi, aifreeapi.com) tracking the OpenAI Sora API, not from a currently live OpenAI pricing page. Omni Flash per-second API pricing has not been published; it is excluded from the pricing row.

The 17× per-second cost spread between the cheapest known SKU (Veo 3.1 Lite at $0.03/sec) and the most expensive (Sora 2 Pro 1080p at $0.50/sec) is the single most important pricing signal in the AI video market as of May 2026. Most launch-week coverage reports headline prices per model without stacking them in a single ranked table. The chart below fills that gap.

Notes on sourcing: Sora 2 API pricing figures come from third-party aggregators (Apiyi and aifreeapi.com) tracking the official OpenAI Sora API — OpenAI's own pricing page no longer hosts a live Sora card post-shutdown. Veo 3.1 pricing is from veo3ai.io and Google DeepMind's Veo model page. Gemini Omni Flash is excluded — no per-second API pricing has been published; the developer API was “coming in the next few weeks” as of May 19, 2026. Omni's effective consumer ceiling is $0 via the free YouTube Shorts surface.

The pricing ladder has a strategic implication that the benchmark conversation misses entirely: the highest-quality Veo 3.1 tier (Quality, $0.20-$0.40/sec) costs less per second than the Sora 2 Pro base 720p tier ($0.30/sec) — and significantly less than Sora 2 Pro's 1080p tier ($0.50/sec). For teams that were choosing Sora 2 Pro for the cinematic quality and not the 1080p output specifically, Veo 3.1 Quality is a credible alternative at lower cost and without the September 24 sunset. This reframes the migration decision for Sora API customers: the question is not just “where do I move?” but “am I overpaying for what I actually need?”

For marketing teams exploring AI video at scale, our AI video automation tools for marketers guide covers the broader workflow integration picture, including how to connect Veo 3.1 outputs to campaign automation pipelines. Our AI transformation advisory can help map the right tier to your specific production volume.

The watermarking and provenance landscape across these three platforms is genuinely novel territory for brand-safety teams — three different mechanisms with three different audit profiles. Most launch-week coverage treats watermarking as a single checkbox. It is not.

SynthID (Google — both Omni Flash and Veo 3.1). Google applies SynthID to every output from both Omni Flash and Veo 3.1. SynthID is a neural watermark — imperceptible to human eyes and ears but detectable by Google's verification tools. It is embedded in the model's generation process and cannot be disabled via the API. Google's Veo model page and Gemini Omni launch post both confirm SynthID as mandatory. SynthID is not the same as C2PA — they are different mechanisms operating at different layers. SynthID is Google's proprietary imperceptible watermark; C2PA is an industry-standard provenance metadata schema.

C2PA + visible Sora mark (OpenAI — Sora 2 and Sora 2 Pro). OpenAI uses two mechanisms for Sora 2 outputs. First, a visible Sora corner mark on downloaded outputs. Second, C2PA provenance metadata — an industry-standard schema supported by the Content Authenticity Initiative (CAI), which includes Adobe, Microsoft, and others. However, OpenAI's C2PA implementation has faced audit concerns: per OpenAI's responsible deployment post, C2PA metadata is embedded on the “no visible watermark” download path. One independent review raised concerns that this implementation is inconsistently applied — frame the C2PA claim as OpenAI's stated policy with those audit concerns noted, not as fully verified.

Brand-safety implication.For enterprise brand teams producing video for regulated industries or high-visibility campaigns, the SynthID approach (imperceptible but verifiable, non-disableable) may be preferable to a visible watermark or an inconsistently applied metadata schema. Veo 3.1 and Omni Flash both carry SynthID — making them more defensible for brand-safety sign-off than Sora 2's C2PA implementation in its current form.

The framework conclusion is that this is not a leaderboard. It is a routing problem. Below is the routing logic for the four most common buyer profiles encountered in video AI decisions in May 2026.

Consumer creator or brand social team (Shorts-first): Use Gemini Omni Flash. The free YouTube Shorts surface is available now. The developer API will follow in the coming weeks. If your workflow is Shorts-native and you were previously paying for Sora or a consumer video AI, Omni Flash is the obvious move. The 10-second output cap is a stated product decision, not a model ceiling.

Creator-tool builder or creative agency: Evaluate Veo 3.1 Quality as your primary path and use the remaining Sora 2 Pro API window for cinematic projects that specifically need longer clips (up to 25 seconds) or where you have existing Sora integrations. Do not build new production systems on the Sora API without a migration plan — September 24 is a hard date. The pricing comparison above shows Veo 3.1 Quality at $0.20-$0.40/sec is cost-competitive with Sora 2 Pro 720p at $0.30/sec. For a broader set of alternatives including Runway Gen-4 and Kling 3, see our Seedance 2 vs Sora vs Kling 3 head-to-head and the Kling 3 at 4K/60fps guide.

Enterprise marketing team (Vertex AI / Workspace): Veo 3.1 is the clear pick. You are already in the Google ecosystem. Veo 3.1 is stable, multi-tier, and has native Workspace integration via Google Vids. The Lite tier at $0.03/sec makes it viable for draft-generation workflows and A/B testing at scale before committing to the Quality tier for hero assets. No sunset risk, no consumer-app drama. Our AI transformation advisory works with enterprise marketing teams on exactly this kind of workflow design.

Team evaluating open-weight alternatives: None of the three platforms in this matrix are open-weight. For teams that need on-premises deployment or full model weights, the relevant comparison is with Seedance 2 from ByteDance and LTX Studio — see our Seedance 2 from ByteDance guide for the open-weight alternative framing. For the Runway GWM-1 universal world model approach — the third major commercial video AI not in this three-way matrix — see our Runway GWM-1 guide.