Google Veo 3.1 vs OpenAI Sora 2: Our Verdict
One model is being actively expanded across three quality tiers on a live API. The other just lost its consumer app and is headed for an API sunset. We tested both to decide which AI video model is worth committing to today.
We recommend Google Veo 3.1 for any team committing to an AI video model in the second half of 2026. Sora 2 still produces the more imaginative footage on stylised and surreal prompts, but OpenAI has discontinued the Sora consumer app and announced an API sunset on September 24, 2026, which takes it off the table for any workflow that needs to exist past the summer.
These two models came out of the same race and now sit on opposite trajectories. Google's Veo 3.1 ships as a three-tier family (Lite, Fast, and Quality) on the Gemini API and Vertex AI, with native audio, 1080p output, and an actively expanding consumer surface through the Gemini app and Flow. OpenAI's Sora 2 launched the most-talked-about consumer video app of the cycle, then in the spring of 2026 lost it: the iOS and web app were discontinued on April 26, 2026, and the API itself is scheduled to stop accepting requests on September 24, 2026.
That's the context any honest comparison has to start from. We tested both on the same brief (short-form social clips, a product walkthrough, a stylised narrative shot, and a tight cinematic plate) and judged them round by round on the work, the access, and the runway. Each round names a winner and states the procedure we used to decide it.
Veo 3.1 produced the more physically accurate footage on every plate involving real-world materials: skin and hair on the café close-up, reflections on the Tokyo wet-streets aerial, and natural shadow behaviour on the field shot. Sora 2 was often the more cinematic composition, but its reflections and contact shadows drifted more often than Veo's. This matches what other testers have reported on the same kinds of prompts.
How we tested itWe generated the same eight prompts in each model at the highest quality tier each offered (Veo 3.1 Quality at 1080p, Sora 2 Pro at 1080p): a golden-retriever-through-a-field shot, a Tokyo aerial at night, a close-up of a face in a café, a slow product rotation, and four scene plates with deliberate physics (reflections, fabric, fluid, caustics). Two reviewers scored each pair blind for texture fidelity, motion plausibility, and lighting accuracy.
On abstract, surreal, and animated prompts, Sora 2 was the more interesting tool. It handled physically impossible scenarios with more grace and held a stylistic identity across regenerations more reliably than Veo 3.1. This is the round Sora wins clearly, and it's the round that explains why Sora's loss from the market matters.
How we tested itWe ran a separate set of six prompts built around the work Sora has historically been praised for: surreal compositions, morphing objects, animated dream sequences, and stylised character shots. We scored on imagination, style consistency across multiple regenerations of the same prompt, and how well each model held a defined look.
Sora 2 Pro supports single generations of 10, 15, or 25 seconds; Veo caps a single generation at roughly 8 seconds, so the same 25-second sequence required three chained Veo generations and a stitch in post. Sora's longer single shots, paired with its Remix, Recut, Blend, Loop, and Storyboard tools, made the narrative round materially easier when we ran it earlier in the cycle on the consumer app.
How we tested itWe listed the supported single-generation durations and the in-product editing tools available to a paying user on each platform's flagship tier, then attempted the same 25-second narrative sequence on both, chaining generations where required.
Veo 3.1 ships native audio across the Fast and Quality tiers on the Gemini API, with the with-audio rate built into the per-second price ($0.15/sec on Fast and $0.40/sec on Standard/Quality). Sora 2 also produces synchronised audio, but Veo's output was more consistently usable as a finished bed in our tests, and audio is now a paid, supported part of Google's rate card rather than a feature tied to a product that's being withdrawn.
How we tested itWe asked each model for the same five clips with native synchronised audio (dialogue, ambient sound, and a music bed under a product shot) and judged whether the audio came out aligned to the picture without a separate pass.
Veo 3.1's three-tier family makes the budget honest: Lite is the draft tier (roughly $0.03–$0.05 per second without audio), Fast is the working tier ($0.15/sec with audio), and Quality is reserved for finals ($0.40/sec with audio). Subscriptions sit underneath that: Google AI Pro at $19.99/month for ~1,000 Flow credits, Ultra at $249.99/month for 25,000. Sora 2's official API is $0.10/sec for the standard 720p tier and $0.30–$0.70/sec for Sora 2 Pro depending on resolution, with ChatGPT Plus at $20/month and Pro at $200/month, but those consumer subscriptions stopped including Sora access on April 26, 2026, so the only remaining official path is the API itself.
How we tested itWe priced a representative month for an independent creator (about 50 finished 8-second clips) and for a small production team (about 500 clips, mix of drafts and finals) on each platform's current rate card, using the cheapest credible tier for drafts and the flagship tier for finals.
This is the round that decided the verdict. OpenAI discontinued the Sora consumer app on April 26, 2026 and has announced that the Sora 2 API will stop accepting requests on September 24, 2026. Google is shipping the opposite direction: Veo 3.1 Lite, Fast, and Quality are all live as preview models on the paid Gemini API tier, Veo runs in Flow and the Gemini app as a consumer surface, and Vertex AI is the enterprise path. A team picking a model today is picking one with a stated end-of-life and one without.
How we tested itWe checked each maker's official status pages and announcements for the live state of the product line: which models are in active development, which surfaces are open, and which endpoints have a published sunset date.
Where the verdict turned
On image quality alone, this would be a closer call. Veo 3.1 Quality is the more physically accurate generator; Sora 2 is the more imaginative one. Spend a week running the same brief through both and you’ll end up with two different highlight reels and two defensible cases for picking either tool.
The verdict isn’t close because of the platform context. OpenAI removed Sora from ChatGPT Plus and Pro subscriptions on April 26, 2026, and the Sora 2 API is scheduled to sunset on September 24, 2026. That’s a hard date attached to the product. Veo 3.1, by contrast, is on its third public revision in twelve months, is split across three live quality tiers on the Gemini API, and is being actively pushed into Google’s consumer products through Flow and Gemini. One of these models is a destination; the other is a sunsetting endpoint.
What each model is actually good at
Veo 3.1’s core strength is photorealistic output. Google’s training pipeline produces video with strong texture detail on natural materials, accurate reflections and shadows, and convincing motion physics. Native synchronised audio is built into the with-audio tiers on the Gemini API, and the model handles standard cinematic shots (close-ups, aerials, product rotations) with fewer regenerations than Sora 2 needed in our tests. Where Veo falls short is on extreme creative prompts: surreal sequences, dream logic, and heavily stylised character work look more pedestrian on Veo than on Sora.
Sora 2’s strengths sit on the other side of that line. On stylised, abstract, and narratively complex prompts, Sora’s outputs are visibly more interesting, and its in-product editing tools (Remix, Recut, Blend, Loop, and Storyboard) were genuinely useful for assembling longer sequences while the consumer app existed. Sora 2 Pro also supports single-generation clip lengths up to 25 seconds against Veo’s 8-second single-generation cap, which made narrative work materially easier. The catch is that the surface where most of that creative tooling lived is now gone, and what remains is an API with a public sunset date.
How to price each one honestly
Veo 3.1’s rate card is the easier one to plan a month around. The API charges per second of generated video and gives you three quality tiers to route between: Lite for drafts, Fast as the working default, and Quality for hero shots. Subscriptions sit underneath the API. Google AI Pro at $19.99/month includes 1,000 Flow credits (enough for roughly 100 Veo 3.1 Lite, 50 Fast, or 10 Quality generations), and Google AI Ultra at $249.99/month opens 25,000 credits. For a working creator running mostly Fast tier at 8 seconds a clip, the Pro subscription is the right starting point; for a small team that needs more than about 200 Fast generations a month, Ultra pays for itself.
Sora 2’s pricing is comparable on paper but harder to commit to in practice. The official API charges $0.10/sec for Sora 2 at 720p and $0.30–$0.70/sec for Sora 2 Pro depending on resolution, and OpenAI’s batch tier discounts both by 50% with a 24-hour SLA. A 10-second standard clip costs about $1.00; a 10-second Pro 1080p clip costs $5.00. The consumer subscriptions are no longer in the picture: ChatGPT Plus and Pro lost Sora access on April 26, 2026, and the API itself stops accepting requests on September 24, 2026.
Who should pick which
Pick Google Veo 3.1 if you’re choosing a video model in mid-2026 and need a stable rate card, native audio, a clear quality tier for drafts and another for finals, and a platform that will exist in 2027. It’s the right default for marketing teams, product video, social-bound content, and anything that has to live inside Google Cloud governance through Vertex AI.
Pick OpenAI Sora 2 only if your work is finishing before September 24, 2026, you need single-generation clips longer than 8 seconds, and your brief leans heavily on the stylised and surreal output that Sora is still better at. Even then, build the project so that the final assets are delivered and accepted before the API sunset, and treat any post-sunset re-edit as a re-shoot on a different model.
For most readers, the answer is Veo 3.1. Sora 2 was the more imaginative tool; it’s also the one being withdrawn from the market.