Sora vs Veo: What Changed in 2026 and What to Use for Product Ads

The short answer: this comparison no longer has two sides. OpenAI announced in March 2026 that it was discontinuing Sora. The Sora web and mobile app shut down in April 2026, and the API is scheduled to end on 24 September 2026. Google's Veo is still live and shipping updates. So if you are choosing between them today, you are not choosing: Veo is the one that exists. The more useful question, and the one this article spends most of its time on, is whether a model like Veo can do the job you actually had in mind, because for product advertising the honest answer is usually no.

Sora vs Veo in 2026: the current state

What happened to Sora?

OpenAI announced the discontinuation on 25 March 2026. The company did not publish a detailed technical rationale, and its public statement was brief, acknowledging that what people had made with Sora mattered. Reporting at the time connected the decision to a broader shift of attention toward enterprise and productivity products, alongside the compute cost of running consumer video generation at scale. The app had also fallen a long way down the App Store charts from its launch peak.

The practical timeline is what matters if you built anything on it. The consumer app and web experience were permanently discontinued on 26 April 2026. The API remains scheduled to switch off on 24 September 2026, which at the time of writing is roughly six weeks away. If you have production workflows still calling it, that is a real deadline rather than a soft deprecation.

This is worth stating plainly because a large number of "best AI video generator 2026" roundups still list Sora near the top. Many of those pages were written in late 2025, lightly refreshed with a new year in the title, and never fact-checked against the product actually existing. If you are researching this category, check the publication and update dates on anything you read, including this article.

What Veo actually does well

Veo generates a short clip, around eight seconds, from a written prompt or a still image, and it generates synchronized audio at the same time. Dialogue, ambient sound and music come out of the same pass rather than being layered on afterwards. That audio capability is the genuinely distinguishing feature against most competitors, where you get a silent clip and then go and solve sound separately.

Access sits inside Google's subscription tiers rather than being sold as a standalone per clip rate. Video generation is included on the Google AI Pro and Ultra plans, with Ultra carrying the higher limits and the higher output ceilings. Google prices these regionally, and when we checked the subscription page it served us euro rates rather than dollars, so confirm the current figure for your own country instead of trusting a number quoted in a blog post. Third parties commonly cite around twenty dollars a month for Pro and around two hundred and fifty for Ultra in the United States, but those are third-party figures and we are not going to present them as Google's published US price.

On raw output quality Veo is genuinely strong. Camera movement is coherent, lighting behaves plausibly, and the audio sync is better than the stitched-together alternative. For imagined footage, mood pieces, concept work and title sequences, it does what it says.

Why neither model can make your product ad

Here is the part that gets buried in every comparison. A prompt to scene model invents every pixel from your text description. It has never seen your product. It cannot have seen your product. So when you prompt it for "a woman holding a matte black water bottle on a kitchen counter," you get a woman holding a matte black water bottle, which resembles yours in the way a stock photo resembles yours. The label is wrong or illegible, the proportions drift, the cap is a different shape, and the next generation changes all of it again.

For a brand campaign where the product is incidental, that is fine. For direct response, where the viewer is being asked to buy the specific item on screen, it is fatal. The person clicking through expects to receive what they just watched.

The second structural problem is consistency. Creative testing depends on changing one variable while everything else holds still, so you can attribute the difference in performance to the thing you changed. A scene model re-rolls the entire world on every generation: new room, new lighting, new actor, new product. Ten variants come back as ten unrelated videos, and you learn nothing about which hook worked. This is not a prompting skill problem. It is how sampling works.

Third is length. Eight seconds is not an ad. A standard thirty second creative needs roughly seventy to seventy-five words of script at natural read pace, which means stitching four or more generations together and hoping they match, which they will not.

What to use instead for product ads

The category you actually want is the one where your text is a script rather than a scene description, and where the tool accepts your real product photo or product URL and composites the actual item into the video. The presenter is generated, the product is yours. We wrote up the full distinction, with pricing for both categories, on the text to video AI page, and the sibling breakdown of starting from a photograph rather than a sentence is on the AI image to video generator page.

In practice a workable stack looks like this. Write the hook first and to a word count, because runtime is arithmetic rather than taste. Feed the product URL rather than a loose image, so the tool picks up the name, the price, the feature claims and the review language along with the photography. Generate several variants that differ in exactly one respect, usually the opening line. Burn the captions in rather than relying on platform auto-captions. Then let spend decide, rather than deciding in advance which one you like.

The raw material for the script is usually already sitting in your store. The phrases customers use unprompted in reviews outperform copywriter language reliably, because they name the objection in the customer's own words, and a tool that collects and organizes that review language turns a scattered pile of feedback into a usable hook list. If you want the mechanics of that specific move, we covered it in how to turn customer reviews into video ads.

None of this means scene models are useless. If you need three seconds of impossible b-roll to cut between talking segments, or an abstract opener, Veo is the right tool and an ad generator is not. Plenty of teams run both. The mistake is expecting the scene model to replace the ad workflow, which is the expensive error this whole category keeps producing.

Frequently asked questions

Is Sora still available in 2026?

No. OpenAI announced the discontinuation on 25 March 2026, the web and mobile app experience was permanently shut down on 26 April 2026, and the API is scheduled to end on 24 September 2026. Any article recommending Sora as a current option has not been updated. If you still have code calling the API, the September date is a hard deadline.

Which is better, Sora or Veo?

Veo, by default, because it is the only one of the two still operating. Before the shutdown the two were genuinely competitive on visual quality, with Veo holding a clear advantage on natively generated audio. That comparison is now historical. For any decision you are making today, the realistic choice is between Veo and the other live scene models such as Runway, Kling, Pika and Luma.

How long are Veo clips?

Around eight seconds per generation, with synchronized audio produced in the same pass. Longer sequences are built by generating several clips and joining them, which introduces continuity problems because each generation re-rolls the scene. If your deliverable is a thirty or sixty second ad, that stitching is the main practical cost.

Can Veo generate a video of my actual product?

No. Veo invents everything in frame from your description, so it produces something that resembles your product category rather than your specific item. Labels, proportions and finish will all be approximations, and they change on every generation. To get your real product on screen you need a tool that ingests your product photo or product URL and composites it into the video.

Can I use Veo output in paid advertising?

Generally yes on paid tiers, but check two things before you spend. First, whether your plan grants commercial use and produces watermark free exports, since free and entry tiers frequently do not. Second, platform disclosure: Meta and TikTok both require realistic AI generated content to be labeled, and both run automated detection alongside the declaration you provide. Labeling costs you nothing. Being caught not labeling costs you distribution.

What replaced Sora?

Nothing replaced it as a single product. The users split across the other scene models, principally Veo, Runway, Kling and Luma, according to whether they cared most about audio, directorial control or cost. For advertisers specifically, the more relevant move was sideways rather than across: to script driven tools that hold the product and presenter constant, because that is the requirement scene models never satisfied in the first place.