Sora vs Veo vs Kling in 2026: one shutdown, one successor, one survivor
One dies in 70 days, one was outranked by its own maker, one still ships. The matchup, re-run for real: six clips, first takes only.
Seventy days from now, on September 24, the Sora API shuts off for good. That is one corner of the most famous matchup in AI video gone from the board entirely. The second corner, Google’s Veo, was just outranked by a newer model from its own maker. The third, Kling, is still shipping and sits mid-table on the quality leaderboard. People keep typing “Sora vs Veo vs Kling” into search boxes anyway, and most of what they find was written when all three names meant something different.
So instead of recycling last season’s verdict, we re-ran the comparison the only way that counts: same prompts, same day, same settings, through the models you can actually buy this week. Six clips, $8.29 in API bills, every result the first thing the model returned. The clips are embedded below so your eyes can disagree with us.
Every name in the matchup has moved
The question froze in late 2025, when Sora 2, Veo 3.1, and Kling were the three names a curious person would compare. The market did not freeze with it. On the Artificial Analysis text-to-video arena (with audio), the current #1 is Gemini Omni Flash at Elo 1,240, a Google model that didn’t exist when the versus query got popular. ByteDance’s Seedance 2.0 sits second at 1,224. The best-placed Kling 3.0 tier is sixth at 1,110, Veo 3.1 is eleventh at 1,093, and Sora 2 is no longer listed at all.
The strangest entry in that list is Google’s own. Gemini Omni Flash reached consumers at I/O in May, opened to developers through the Gemini API on June 30, and Google says outright that Omni will replace Veo in the Gemini app. Veo 3.1 remains on sale as the high-fidelity option, but when a vendor starts describing its former flagship as the legacy pipeline, you should hear a clock ticking. Google is beating Veo so you don’t have to.
Kling moved too, in the other direction. Kuaishou’s February 5 launch of Kling 3.0 added native audio in five languages (English, Chinese, Japanese, Korean, and Spanish, plus accents and dialects), clips up to 15 seconds, and a multi-shot storyboard mode in its Omni tier that lets you specify shot size, perspective, and camera movement per shot. It stopped competing on “prettiest five seconds” and started competing on “most of a scene.”
None of this is unusual for the category; it is the category. Video models turn over faster than any other modality right now, fast enough that our spring roundup logged three new flagships and one corporate exit in a single quarter. Any versus article, including this one, is a photograph of a moving object. That is why everything below is dated, priced, and reproducible for a few dollars rather than argued from authority.
Sora is a deadline now, not a contender
The short version of the Sora story: OpenAI shut down the app and website on April 26, and the reporting pointed at economics, roughly a million dollars a day of inference against a shrinking user base. We covered that Sunday, and what it says about consumer AI video, in our spring video roundup. What matters now is the second date: OpenAI’s deprecations page lists sora-2, sora-2-pro, and the entire Videos API for removal on September 24, 2026, and names no replacement.
If you have a Sora integration in production, the versus question is not academic: you have ten weeks to pick a lane, re-test your prompts, and move before the endpoint starts returning errors. Budget for surprises beyond the model swap itself. Durations, aspect-ratio options, and content policies differ between vendors, and a prompt tuned for one model’s style reads differently to another. If you were just deciding what to try this weekend, cross the name off entirely. A model you cannot buy is not a contender, whatever its demo reel looked like. That leaves the actual July 2026 head-to-head: Kling 3.0, Veo 3.1, and the model Google sent to replace Veo.
The juggler test: one sentence, three different films
Our first prompt was built to stress motion, physics, and camera direction at once: a street performer in a red coat juggling three flaming torches at dusk, slow dolly-in from a low angle, crowd in soft focus, flames reflecting on rain-wet cobblestones, ambient sound. One sentence, five constraints. We sent it to all three models at 1280×720, six seconds, audio on, through Runware’s API, and kept the first result. No seed pinning this time: two of the three reject a seed parameter, so every clip is an honest single roll.
Three models, one sentence, three different films. Veo shot the poster: the best light, the best crowd, the best plaza, and the loosest grip on the instruction, multiplying three torches into a small fire show that ends uncomfortably close to the performer’s face. Kling shot the floor: it obeyed the low angle past the point of usefulness and held a static camera, but its surfaces are the most photoreal in this post. Omni Flash shot the coverage: subject framed, camera moving as directed, everything a director could cut around, rendered in the flattest light with the least convincing fire.
No clip shows exactly three torches cleanly in play for all six seconds, which is worth saying plainly: prompt adherence on counts and physics is still the weakest axis in AI video, at every price. If your shot depends on a number staying true, plan to roll more than once.
The clock matters as much as the frame. Omni Flash returned its clips in 33 and 34 seconds; Kling took 50 and 81; Veo took 72 and 92. On a single clip that difference is trivia. Across an afternoon of iteration it is the difference between trying eight ideas and trying three, and iteration count decides quality more often than model choice does.
The dialogue test: all three said the line, word for word
The second prompt tested speech: a gray-haired barista behind an espresso bar looks into the camera and says, “We close at nine, but the machine sleeps at eight.” A specific line is a harder test than ambience because you can grade it exactly. We ran each clip’s audio through Whisper afterward (a fraction of a cent per clip on Replicate), and all three transcripts came back verbatim.
Eighteen months ago, a talking character with synchronized audio was a separate product category. Today it is table stakes: the cheapest model in this test delivered scripted dialogue as reliably as the most expensive one. When every contender passes your test, the test stops being the decision. The decision moves to staging taste and, mostly, to price.
Be clear about what one line in English does not prove. We did not test Kling’s other four dialogue languages, fifteen-second multi-shot scenes, or how any of these models hold a face steady across cuts, and character consistency remains the open problem in text-to-video. A wedding videographer and a TikTok ads team should not read this test the same way. It proves the floor is high now; it says nothing about your ceiling.
Price per second spans 8x, and quality doesn’t track it
Here is the axis the showreels never show. One second of 720p video with audio, priced across the current field, using what we were actually billed where we ran the model ourselves and Google’s published prices where we didn’t.
Read the table against the leaderboard and the discomfort is obvious: the arena’s #1 model costs a quarter of #11. Veo 3.1 produced our most beautiful clip and billed $0.40 a second for it, while the model Google built to replace it billed $0.10 and came back twice as fast. Kling’s $0.126 buys the best surface detail in this post, and its tiers stretch from about $0.42 a second for the 4K mode down to budget rates without audio. Sora 2’s $0.75 now reads like an artifact from another era, which is exactly what it is. The cheapest listed rate on the board, Veo 3.1 Lite’s $0.05, is 8x below its own family flagship. We priced this collapse across modalities in cost-per-task, not cost-per-token; video is where it moves fastest.
Multiply by sixty to feel the stakes. A minute of finished footage costs about $3 on Veo 3.1 Lite, $6 on Omni Flash, $7.56 on Kling Standard, and $24 on Veo 3.1. Nobody ships every generated second, so real budgets run several times those figures once you account for rerolls. At those multiples, knowing which cheap tier passes your tests is worth real money every week.
Route by job, not by brand
Put the clips and the prices together and no single winner survives. What survives is a routing table.
Two entries deserve a word. First, Seedance 2.0 appears in the dialogue row because the arena ranks it second overall with audio, ahead of everything we tested here; we put it through the same first-result treatment in our ByteDance deep-dive, and it belongs on any 2026 shortlist even though nobody types its name into versus queries yet. Second, the Sora migration row favors the Veo family not because Google won the bakeoff but because the request shape matches: short clips, native audio, image conditioning. Migration is about changing the fewest assumptions, then re-evaluating once you are stable.
Every model in this post has a page in the CSuite catalog with pricing and capabilities: Kling 3.0, Veo 3.1, and Gemini Omni Flash, all reachable at provider cost through your own Runware or Replicate keys in the app; the full lineup is at /models.
Retire the versus search; run your own
The uncomfortable lesson of this post is not that one model won. It is that the question aged out from under the answer in about two quarters. Today’s podium holds a Google model nobody was searching for in March and a ByteDance model most people still can’t name. Whatever verdict we hand you today has the same half-life.
So take the method instead of the verdict. Write two or three prompts from your real workload: one with motion you care about, one with a line of dialogue, one with a constraint that must hold. Run them through the two or three models your budget shortlists, first result only, and keep the receipts. Ours came to $8.29, most of it spent learning that the expensive model is gorgeous and disobedient, the cheap one is fast and literal, and the mid-priced one quietly does the job. That is more than any leaderboard told us. The template for turning this into a repeatable habit is in our guide to writing your own eval.
And date-stamp everything, including us. These numbers were checked on July 16, 2026. If you are reading this in the fall, Kling will have a new tier, Google will have folded Veo deeper into Omni, and someone else’s name will be missing from the board. Rerun the clips before you believe anyone, including the version of us that wrote this.
