Veo 3

By Minel Gunesoglu, founder. I build the score card and read the launch-week community threads myself before any number ships. Prices collected 9 July 2026, the arena row taken from a 17 July 2026 capture of the board, and the card re-run on 19 August 2026. I revisit this page monthly; one of its four rows measures access terms rather than the axis it sits on, so read that row's own heading before reading the 79.0.

Two products answer to the name Veo 3. This review is about Google's Veo 3 video model, the Google DeepMind system you reach inside the Gemini app to turn a text prompt into an eight-second clip with sound, and not the Veo Cam 3 sports-tracking camera that Veo Technologies sells to soccer clubs. With that settled: Google's Veo 3 composites 79.0 out of 100 on this desk's card, across four filled rows, one of which is not the axis it occupies. The score leans heavily on how cheap the model got in June, not on how far ahead it sits of its rivals.

Two things pull against each other here. On the Artificial Analysis Video Arena, Google's Veo 3.1 sits at an ELO of 1,088, eleventh where audio counts toward the comparison and inside a nine-to-fourteen tie band, in a launch cycle that treated it as the clear frontier. At the same time, the June price cut made it one of the cheapest ways to generate video with native audio at all. A review that tells you only one of those facts is selling you something.

Veo 3 pricing: what AI Plus, Pro and Ultra cost, and whether Veo 3 is free

Start with the number that decides most purchases, because Google moved it recently and much of the web has not caught up. There is no standing free tier for Veo video: Google's own subscriptions page shows the free Gemini plan does not include Veo video generation, and the occasionally cited roughly ten free videos a month via Google Vids claim is weakly sourced, so I leave it out rather than repeat it. It is set in italics rather than quotation marks because it is a claim in circulation, not anybody's quoted words. Paid access begins with Google AI Plus at $4.99 a month, cut from $7.99 on 8 June 2026, a drop 9to5Google logged the day it landed. That tier carries 200 Google Flow credits a month. Above it, Google AI Pro runs $19.99 a month, and Ultra is sold from $99.99 a month: the plan card prints From $99.99/mo and the comparison table underneath it names two separate Ultra tiers, Ultra 5x at $99.99 and Ultra 20x at $199.99. Prices read from Google's own US plans page on 12 August 2026.

The credit allotment is published, and it is the row worth reading before the price. Google's plans page carries a Google Flow Credits line reading 200 a month on Plus, 1,000 on Pro, and 10,000 and 25,000 on the two Ultra tiers. What that page does not do is connect any of it to Veo: its Video generation model access row reads Omni Flash on all four plans, and the string Veo does not appear in the document at all. So the credits are sourced and the model mapping is not, and this page states the first and refuses the second rather than filling the gap from a press summary.

That $4.99 figure is the single most mis-stated fact about this model. Recent third-party Veo 3 reviews still quote the pre-cut $7.99 price, invent tier names and figures that do not match Google's actual plans at all, or drop the cheap tier entirely and frame Veo 3 as an Ultra-only product. This page carried a fourth error of its own until 12 August 2026: it priced Ultra at a single figure Google's plans page does not print anywhere, which is the same mistake read from the other side. Each of those errors makes the model look more expensive or less accessible than it is. The price in the table is dated and sourced; re-confirm it on Google's plan page the day you buy, because on this product the number keeps moving.

Two routes soften the no-free-tier wall. Google runs a 30-day trial on the Pro tier, and it offers a year of Pro free to verified students, so "is Veo 3 free" has a real, if temporary, yes attached to it. For access itself, consumers reach the model in the Gemini app and in Google's Flow editing surface; businesses reach the same family through Vertex AI Studio, a different, more enterprise-shaped path that recent reviews describe as harder to navigate, and one that has carried a reported $300 / 30-day free-credit offer I have not re-verified this quarter.

Whatever tier you land on, budget in credits and clip length, not in the monthly dollar figure. Every Veo 3 generation caps at eight seconds, on every plan, and the credit meter draws down fast enough that a short experiment can empty an allotment before you have a keeper. Anything longer than eight seconds means stitching clips together in Flow's extend-and-sequence workflow rather than generating one long take. Which model versions actually support that extension is reported inconsistently across recent write-ups: some tie it to Veo 3.1, others say clip-extension still only runs on the older Veo 2. Treat the exact version support as unsettled until Google documents it plainly. The eight-second per-generation ceiling itself is not in dispute.

What is the Veo 3 generation limit on Google AI Ultra?

Ultra is sold on the promise of higher limits, and that promise is the problem, because Google's marketing language does not map to a number a subscriber can find in the product. In Google's own Gemini Apps Community, paying Ultra customers open thread after thread trying to establish what their actual Veo 3 allowance is. "Lack of Clarity on Veo 3 Video Generation Limits for AI Ultra Plan" (Gemini Apps Community, 2025). "4 video Veo3 daily limit on Ultra?" (Gemini Apps Community, 2025). "Video generation limit is ridiculous" (Gemini Apps Community, 2025). These are separately opened threads, not one loud outlier. So I will not print an Ultra generation cap here, because there is no documented number to print. The honest statement is that the most expensive tier is also the one whose real ceiling Google leaves opaque.

How good is Google Veo 3? The score card and its Video Arena rank

Here is the whole card, and the first thing to notice is which row is not what its heading would normally mean. Value scores 92.0, on the strength of that $4.99 entry against what the model can do, and it is the figure holding the composite up. Capability scores 69.4, drawn straight from the Video Arena ELO and graded as thin evidence, because the vote count printed beside that row never entered the scoring input. Commercial terms reads 81.2 off two positions in Google's own paperwork. The usability row is the odd one: no consumer record of this product exists that can honestly be attributed to it, so since 19 August 2026 that row publishes access terms: 75.2, a reading of what Google puts in writing about getting and keeping access. So the composite is 79.0 out of 100 (4.0/5), prices collected 9 July 2026 and the arena row read off a board archived on 17 July.

Read what the access-terms row is before reading the total. It measures Google's published commitments, not anybody's experience of the product, and one of its three readings needs a caveat that the rubric's own label does not carry: Google One publishes chat, phone and email support, and the reading scores 70.0 because those channels are published once for every member rather than split by tier. Nothing here says Google offers one support channel. The regional reading works the same way, a closed allowlist of 168 countries and regions is the more transparent disclosure and the more restrictive product, and this position measures availability, so it scores below a published list of exclusions rather than above one.

One thing about the movement of this number deserves saying out loud rather than leaving in a methodology note, because it has now gone both ways for the same kind of reason. The composite rose on 12 August 2026, from 76.3 to 80.5, and no measurement of this product improved: the cell that disappeared was the lowest of the three, the composite is a geometric mean over the rows that carry a value, and removing a below-average row lifts a mean. The arena correction later the same day pushed three tenths back off it. Then on 19 August 2026 it fell, from 80.3 to 79.0, and again nothing about Veo 3 changed. The site stopped publishing empty rows, the last blank took an access-terms reading of 75.2, and a figure below the running total drags a geometric mean down. That fall crosses a band boundary: this card read high until 19 August and reads above average now, on a one-and-a-third-point move. A band edge is an arbitrary line drawn through a continuous scale, and it is worth knowing that a card sitting near one can change its word without changing its evidence.

One number that could have been on this page is deliberately not, and why it is missing is worth more than the figure would have been. A browser read on 12 August 2026 found a Trustpilot profile for Google's Gemini domain at 1.5 out of 5 across 1,078 reviews. Nothing on this card reads it, because that profile belongs to the Gemini assistant and not to Veo 3. Veo is a video model you reach through Gemini, Flow and Vertex, and the people reviewing that domain are overwhelmingly writing about a chat assistant and what it charges them. Counting them here would import a different product's ratings, which is precisely the error this desk made on 11 August 2026, when a shift-planning company listed as Flair was counted as the incumbent for an image tool of the same name. The narrower labs.google profile was read the same day and carries four reviews, under the ten this scoring will look at.

The consequence of that ruling has to be said out loud, because it runs in this product's favour and it was re-examined on 19 August 2026 rather than inherited. A rating that low on a sample that large would enter the usability row as arithmetic and drag this composite down by more than fifteen points. The reading that filled the row instead cost it 1.3. So the identity refusal is worth an order of magnitude more to the number at the top of this page than the criterion that replaced it, and it was still refused, on what the reviews are about rather than on what they would cost: people writing about the Gemini domain are writing about a chat assistant, and what people who use Gemini report is not what people who use Veo report. The record this replaced could not be defended either way. It published a sample size of 891 with no rating attached to it, and 891 is findable in no capture on file.

On raw output, the arena result is the honest ceiling on what I can claim. Google's Veo 3.1 earns an ELO of 1,088 on the Artificial Analysis Video Arena for text-to-video, eleventh on that board's audio-inclusive ladder as archived on 17 July 2026. The figure this page carried until 12 August was 1095, recorded as a 9 July reading, and no archive of the board exists for 9 July, so the number moved to the nearest day that can actually be opened rather than staying on one that cannot. Eleventh is respectable, and it is not the frontier, which is the single most useful correction to the launch coverage: this is a good model, not the best-scoring one on an independent board. None of the competing Veo 3 reviews I read cite any independent leaderboard; the only benchmark most of them quote is Google's own internal preference number. If you want the exact rule that turns an arena ELO into a capability score, how the arena rank is built is written out in full. What I have not done is run my own controlled generation tests, so I am not going to tell you Veo 3's real-world failure rate or how long a render takes under load. Those questions need a measured run I have not made, and no cell on this card claims to answer them.

One completeness note on the legal side, because tool reviews tend to skip it, and it matters more now that no cell carries it. Content moderation applies on the hosted product, and Google does publish a generative-AI indemnity, but read where it lands. The published cover reaches Veo through the Gemini Enterprise Agent Platform API, formerly Vertex AI, and through Google Cloud. The $4.99 consumer plan this page prices is a different door into the same model, and the indemnity is not written to it. Scoped rather than absent is the accurate description, and the scope is the part a buyer taking client work needs before anything ships.

Does Veo 3 generate audio? Native sound, and the "no sound" reports

The feature that sold Veo 3 was not its picture, it was its sound. On both Hacker News and Product Hunt, the loudest launch-week reaction was surprise that the model generates synchronized dialogue and sound effects in a single pass, instead of handing you a silent clip to dub later. "Blown away at how well the audio/video matches up, and the dialogue is better sounding / on-par with dedicated voice models," one commenter wrote on Hacker News (20 May 2025). A commenter on the Product Hunt launch thread put the promise plainly: "The integration of audio generation with video will open up so many creative possibilities" (Paul Hua, Product Hunt launch thread). Until 12 August 2026 this page called him a Product Hunt reviewer. He is not one: the product's Reviews section carries no reviews at all and still invites the reader to be the first, and the same is true of the second Product Hunt voice quoted further down. Both are launch-thread comments, which is a different kind of evidence and a weaker one, and it matches the score card, where no cell reads a consumer rating for this product. Product Hunt prints no absolute date on these comments. Every one of them read "1yr ago" on 12 August 2026, against a May 2025 launch.

Then, within about a week, the same feature broke for a subset of users. Google's Gemini Apps Community began filling with reports of clips rendering completely silent despite an explicit audio prompt, alongside a second bug where a generation-limit message kept pushing its own reset time forward. The poster quotes the message as "You've reached your video generation limit until May 20, 16:24 AM", says a refresh a minute later returned the same error with the time moved on again, and draws the conclusion as a conditional rather than a prediction: "At this rate, I will NEVER be able to create another video again, as the limit keeps resetting to 24 hours later." (Gemini Apps Community, Darren wall, 19 May 2025; the thread is locked and carries 32 replies.) Until 12 August 2026 this page welded the error string and the never-again clause into one quotation with an ellipsis, which dropped both the opening qualifier and the cause the poster gave. A separate, recurring complaint is that on-screen text and captions generate garbled or misspelled, which matters the moment you want readable words in a clip. Native audio is genuinely the standout capability. It is also the one with the best-documented launch-week reliability problem, and both facts belong on the same page.

A version note belongs with the audio story, because most of what you generate now is Veo 3.1, not the original 3.0. AI Plus ships Veo 3.1 Fast, the higher tiers ship full Veo 3.1, and the arena figure above is measured on 3.1. Google positions 3.1 as an iteration that tightened audio-video synchronization and general stability over the launch build, which is the version most current reviews test against. Beyond that release framing, Google does not publish the underlying audio specifications (sample rate, encoding format, or sync latency) in a form I can independently verify, so I will not attach numbers to the 3.0-to-3.1 audio gap; read it as a real but unquantified improvement until Google documents the spec.

Veo 3 limitations users hit: strict moderation and country availability

Two more patterns show up consistently enough to report, both from named, dated threads rather than vibes. The first is a control-versus-convenience split between audiences. Product Hunt's launch comments run uniformly celebratory, while Hacker News's more technical crowd, reacting to the identical announcement, worried that a hosted tool is not production-ready precisely because of platform-side moderation. "you're not limited by the platform moderation that can be too strict and arbitrary and fail with the false positives," one practitioner argued on Hacker News (20 May 2025), comparing Veo and Flow against open, locally run models. Read that as a real trade-off, not a knock: you buy convenience and give up some control over what the moderation filter will and will not pass.

The second is geography. Veo 3 shipped US-first, and the frustration landed on two groups at once. Onlookers noted that the demo was not available where they lived: "Note that it's not yet available in all countries" (Sean Tiffonnet, commenting on the Product Hunt launch thread; the surface carries no absolute date, and the launch is May 2025). More sharply, paying Ultra subscribers abroad reported buying the top tier and still not seeing the model: "i paid for the ultrs package, am in the uk and can't get veo 3. it only shows 2.5 pro as the option" (Gemini Apps Community, 2025). Paying full freight and not getting the product is the version of the pricing story that actually stings.

One honest limit on all of this: the community record for Veo 3 is thin and skews to the May 2025 launch window. It is real, and it is dated, but it is early-days evidence (quotes rather than a rated, counted sample), which is why the usability row on the card publishes access terms instead of anything drawn from customers.

Veo 3 as a Sora alternative after the Sora shutdown

A lot of the people arriving at Veo 3 right now are former Sora users looking for a new home. OpenAI announced on 24 March 2026 that it was pulling the standalone Sora consumer app, and the app went offline on 26 April 2026, which sent a wave of creators shopping for a replacement. Veo 3 is one of the names that comes up most, and the arena rank and the $4.99 entry above are most of why. This is not a migration guide, and there is no Sora-against-Veo-3 page coming: that format went on 11 August 2026, and the promise of one is withdrawn here. Sora is also no longer purchasable, which makes a matchup between them a comparison with a product a reader cannot buy. The short version: the switch is common, the price of entry is low, and the eight-second cap and the credit math are the two things that will surprise a former Sora user first.

Is Veo 3 worth it? Alternatives in the category, and the verdict

The video-generators hub ranks every tool in this category on one scale and one rubric, which is what lets you set Veo 3's 79.0 against the rest of the field without mentally converting between one reviewer's numbers and another's. One qualifier belongs on that, and it is new as of 19 August 2026: where a card could not fill a row, the row carries a named criterion instead of the axis, and the card says which. Two cards in this category do that, this one included, so read the row headings before reading a gap between two totals. If Veo 3's eight-second ceiling or its opaque Ultra limits are dealbreakers for you, the hub is still the place to weigh the alternatives.

The verdict, dated 19 August 2026: Google's Veo 3 is a good, genuinely affordable video model that the launch cycle oversold. Its independent arena placing is eleventh, its real strength is one-pass native audio, and its 79.0 composite is propped up more by a $4.99 price than by best-in-class output, with one of the four rows behind it reading Google's published access terms rather than anything a customer reported. If you make short clips and want sound without a second tool, the entry price makes it easy to try. If you are budgeting real production time, price in the eight-second cap, the fast-burning credits, the launch-era reliability reports, and an Ultra tier whose ceiling Google will not put a number on. Strip the launch-week noise away and what is left is an eleventh-placed video model that happens to be cheap and can talk: genuinely useful, a long way from the frontier the hype sold, and priced like Google knows the gap.

Scored and written by Minel Gunesoglu, founder of Vouch. Find me on LinkedIn, or read how the audit works. Every price, rank, and rating on this page is dated where it appears and re-checked monthly. Disclosure: Google runs no affiliate program for these plans, so the links to Google's own pages here are plain, unpaid links. Where other outbound links on this site are affiliate links they are marked, and paid submissions only ever buy a faster review, never a listing or a better score, which are fixed by the methodology alone.