If Veo 3 is the Regina George of AI video, then Sora is Cady Heron. That’s the conclusion I’ve come to after rigorously testing the popular AI video generators.
Google’s Veo 3 was the first AI video model to introduce synchronized audio, and it raised a lot of fanfare for this latest wave of generative AI tech. OpenAI’s Sora was the new girl in town who met with fervor and adoration, with something new to offer. Like the iconic Mean Girls movie, both reveal ugly but fundamental truths about where we as a society are headed as AI slop fills our feeds.
During the initial stages of my testing, Sora 2 and Veo 3 appeared to be evenly matched. They have a lot of the same capabilities and features. Fundamentally, they each generate realistic AI video clips with sound. One is free for a limited time; one starts at $20 per month. But the free one is slower than the model behind the paywall.
I’ve used and reviewed a lot of AI image and video generators, but trying to crown a winner was so frustrating — especially amid very real, significant concerns that the rise of AI video technology is leading to more accessible deepfakes and making it harder to tell what is real versus AI.
After creating a slew of AI videos, an ultimate winner emerged. One service is just a bit better. The videos flow more smoothly and have cleaner, aptly fitting audio. Both are excellent at what they claim to do. But there can be only one queen bee at North Shore High School. Here’s how our testing of the new Sora 2 and Veo 3 AI video models went.
Sora overview
Sora is OpenAI’s video generator. You probably know OpenAI as the company that makes the uber-popular chatbot ChatGPT. Sora is also the name of OpenAI’s TikTok-like social media app. For the purposes of this evaluation, we’re comparing the Sora 2 model, not the social app, to Google’s Veo 3 model. The AI social media app is unique, but there are many concerns around its ability to quickly and easily produce deepfakes and misinformation.
Sora launched in December 2024, and the new Sora 2 model brings much-needed upgrades. Sora videos can be 10 to 15 seconds long, with audio (with Sora 2) up to 1080p. You can use Sora for free and no longer need an invite code. You can unlock more features with paid ChatGPT plans.
To identify an AI-generated clip, Sora videos have C2PA metadata and a visible, cloud-shaped, bouncing watermark. OpenAI’s privacy policy allows you to turn off AI training on your content, and its approach to image and video generation outlines how it attempts to prevent the creation of harmful and inappropriate content. (Disclosure: Ziff Davis, CNET’s parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
Veo 3 overview
It wasn’t until the third generation of Veo that Google’s AI video generator really took off. My disappointing experience with Veo 2 was quickly remedied. The most important feature Veo 3 introduced was also a first for a big tech/AI company: AI videos with synchronized, AI-generated audio. It might not seem like a big deal now, but it was game-changing when Google dropped the new model at its 2025 I/O conference.
Veo 3 videos are 8 seconds long at 720p in the Gemini app (though those specs can be changed in different Google AI tools, like its AI filmmaking program, Flow). Veo 3 videos have a visible «Veo» watermark in the lower right corner (unless you want to cough up $250 for the Ultra plan to remove it) and an invisible SynthID watermark that’s embedded in the metadata.
Google’s Gemini privacy policy says the company can use your personal info to improve its technologies, which is why it recommends not sharing sensitive or confidential information with it. Google’s generative AI prohibitive use policy seeks to prevent the creation of abusive and illegal content.
Don’t miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.
Video and audio quality
Sora lets you make longer videos (10 to 15 seconds, 25 for Pro users) and choose between portrait and landscape orientations before you generate. Veo videos are always 8 seconds long and landscape — you can’t change this, and attempts to do so in your prompt likely will be unsuccessful. (Developers, however, can adjust Veo’s aspect ratios.)
Beyond each video’s specs, what distinguishes a good AI video is a simple test: Can it obey our planet’s laws of physics? When I asked each service to generate a ballerina dancing, I wanted her to effortlessly glide across the floor — not float 3 feet above it, for example. Sora and Veo were once again matched here. Sora’s ballerina was much more lifelike, but Veo’s alien didn’t sprout new arms and legs during the dance battle clip. It’s a great example of how even the so-called best programs will flub or hallucinate at some point.
Both programs synchronized audio well, though I was happier with Sora’s audio overall. It often added the appropriate music or background/white noise without me asking. For example, Sora added classical music for my ballerina and a conversation between a barista and a customer picking up a latte out of frame in my cafe video. Veo’s audio is good, and it was the first to add audio to its capabilities, but I have to give the edge to Sora now for going the extra mile.
Prompt adherence and speed
Sora and Veo 3 had good prompt adherence, meaning they created the video I asked for. One of the best ways to test prompt adherence is to ask it to include specific text or audio. Veo 3 was better at creating clear text, as you can see in the videos of «Katelyn’s Cafe.» I’ve had my name misspelled worse but I was disappointed in Sora’s hallucination there.
Another way to measure prompt adherence is by using more complex prompts. Your prompts don’t need to be over the top, but previous AI models have struggled with negative or exclusionary characteristics — things you don’t want it to include. I was pleased that both excelled at negative directions, like «don’t change the plate’s floral print.» Upgrades from previous models are likely to be thanked for this perk.
Prompt adherence is important because neither program has good ways to edit your videos after they’ve been made. Sora’s «edit video» button in your drafts folder will only allow you to rewrite your prompt and regenerate. Follow-up requests to Gemini don’t always work. It’s frustrating, but I expect both companies will focus on it in the future.
Veo generated its videos quicker, even if just by a minute. But both had average video generation speeds between two and five minutes.
Pricing
While Sora 2 is free for a limited time after its launch, it does not guarantee that it will last forever. You’ll need an invite code to use Sora, either through the social app or web.
You can’t use Veo 3 for free; the cheapest way to use it is with Google’s $20 per month AI Pro plan. Veo 3 is also available through Vertex AI, the Gemini API and Flow. If you don’t know what these developer-focused tools are, you likely don’t have access to them.
You may need to upgrade your Google plan from there if you need to generate a lot of videos; generation limits increase with each tier. I hit my limit after five videos, locking me out for 4 hours. I recommend starting with the lowest possible plan and upgrading once you’ve played around with the tool and understand how it will fit into your workflow.
Sora is clearly the winner in terms of price, but OpenAI will likely make us pay up to use the new model eventually. Even then, OpenAI’s ChatGPT Plus tier is also $20, the same as the lowest Google AI plan. From there, it will depend on how the generation limits differ. OpenAI did not respond to a request for comment on how long Sora will be free by the time of publication.
Verdict: In a decaying reality, Sora has a slim edge
I’ve been reviewing generative AI creative software for nearly as long as it’s been around. But there was something particularly horrifying about watching Veo 3 and Sora run side by side and pop out ultrarealistic videos in under 5 minutes.
AI video is no longer a niche fascination but a part of our new reality. I’m an AI expert, but I wouldn’t be able to tell if some of these videos were AI-generated without the watermarks. It’s an unsettling feeling, to say the least, and testing Sora and Veo reminded me of why it’s so important to label AI content.
Sora and Veo are impressive programs that are likely to fit your needs. If you already pay for ChatGPT or Gemini, there’s no need to branch out to the other. Each company’s respective AI video generator is going to work for the majority of projects.
I have to give the ultimate edge to Sora. Its generations have smoother motions, fitting audio and fewer hallucinations. The generations might be a few seconds behind Gemini, but those results are going to be worth the wait. Its bouncing watermark, unlike Veo’s static one, makes me feel a bit better about its ability to create free, realistic deepfakes. From a technical standpoint, Sora is the current reigning champ.
I can’t in good conscious leave out the fact that both of these AI video models make it easier to confuse reality with AI and struggle to stop the creation of abusive content. Sora’s chief feature in its social media app is the ability to use your and other people’s likenesses in nearly any kind of AI video. Celebrities like Breaking Bad actor Bryan Cranston, talent agencies and workers’ unions have pushed OpenAI for stricter Sora guardrails, which it has delivered. But those guardrails aren’t perfect. Cranston’s concerns came days after OpenAI had to temporarily pause generations of Martin Luther King Jr., at the request of his estate, after weird and racist deepfakes of the civil rights leader flooded the app. Google has struggled with similar issues with Veo 3 and its AI image models.
The AI industry is straining to create and implement effective policies and technical guardrails to protect copyright holders, celebrities and public figures — and all of us who see these videos online. These two titans aren’t immune to those challenges.
Also worth noting is that AI video is a rapidly evolving field. A new update to either model could make its usefulness skyrocket or plummet. Sora’s limited-time free access is a perk now, but Veo could become the better choice overall if OpenAI puts Sora behind a paywall with its more expensive subscriptions.
When to use Sora: What Sora is best at
Sora is best at photorealistic videography and would be best suited for amateur creators and AI enthusiasts. Sora has more in-app settings you can toggle, like changing the orientation and length of videos. Sora videos might take a minute or two longer to cook, but you won’t run into generation limits as quickly as you will with Gemini. And for better or worse, Sora can generate realistic products like iPhones and some celebrities and public figures with no trouble. While that’s not a good thing for us as a society, it is a sign of its AI prowess.
When to use Veo 3: What Veo 3 is best at
Veo 3 is better for professional-minded creators. Veo’s creativity and prompt adherence are excellent, and its conversational flow will feel familiar to a typical chatbot experience. My Veo videos also feel less intense than Sora videos. The animated Veo donut I created was much more tame and cute than my mildly unhinged Sora donut, for example. I also appreciate that you can use Veo in Google’s more professional AI programs, like Flow, that give you more hands-on editing tools than you get just in Gemini.
How we tested Sora vs Veo 3
To make things as fair as possible, I gave each AI the exact prompt for each test. I used a variety of scenarios to see where each excelled: animating my existing photos, complex sci-fi dance battles and ultrarealistic stock videography, to name a few. I tested them through their web browsers and mobile apps.
For more, check out why it’s so hard to tell AI videos from real videos.
