AI Video Generation in 2026: Kling, Sora, Veo and the State of Text-to-Video Models

Last Updated: October 2026

When this site first covered Kling in 2024, the story was a challenger: a Chinese short-video company had shown clips that looked like OpenAI’s unreleased Sora. Two years later the premise has inverted. Sora’s consumer app closed on 26 April 2026 and its API went offline on 24 September 2026, while Kling, Google’s Veo and Alibaba’s open Wan family became the tools people actually build on.

That shift matters because AI video generation is no longer a demo category. It has per-second price lists, API deprecation schedules, provenance rules and a visible gap between marketing reels and what a production pipeline can rely on. A model you integrate today can be switched off, repriced or quietly replaced within a year, as the Sora API’s fate shows.

You will leave with a working mental model of how modern text-to-video systems are built, a verified snapshot of the main players, an honest cost calculation, and a selection and quality-control workflow you can adapt.

What this covers: the market after Sora, the diffusion-transformer architecture with a worked token-count example, a comparison of Kling, Veo and open models, cost maths, evaluation with VBench, failure modes, provenance, and a practical checklist.

Context and Background

Text-to-video means mapping a natural-language prompt, optionally with reference images or audio, to a few seconds of moving pictures. The field spent 2022 and 2023 on U-Net diffusion extensions such as Stable Video Diffusion and AnimateDiff, which bolted temporal layers onto image models. Results flickered, clips were short and identity drifted within a second or two.

The turn came with transformer backbones trained on compressed spacetime patches. OpenAI’s Sora technical report popularised the recipe in early 2024, and Kuaishou’s Kling, first shown in June 2024, was the first widely seen system to match the look of those demos. Since then the lineage has been fast. According to the Wikipedia history of Kling, version 1.6 arrived in December 2024, 2.0 in April 2025, 2.1 in May 2025 and 3.0 as a stable release on 7 February 2026. Google’s Veo went from a May 2024 announcement to Veo 3 in May 2025, which added synchronized audio, and Veo 3.1 in October 2025.

The incumbents list changed in 2026. OpenAI announced Sora’s discontinuation in late March 2026, with the web and app shutting on 26 April and the API following on 24 September, according to The Decoder’s reporting. OpenAI’s own model page now states that the Sora 2 models and Videos API were shut down on 24 September 2026 and that there is no one-to-one replacement. Press coverage frames this as a reallocation of compute toward coding and enterprise products, with Sora continuing as world-model research; OpenAI itself has not, in what I could verify, published a detailed rationale.

For readers tracking the wider frontier-model landscape, our fact-check of viral AGI videos shows why provenance matters as realism improves, and our earlier piece on Kling O1 and unified video editing covers one branch of the Kling family in more depth. The architectural idea of routing compute through specialised sub-networks, which one open video model now uses, is explained in our mixture-of-experts guide.

My thesis for this post is simple. The technical story of 2024 was “can anyone match Sora”. The 2026 story is vendor risk: the models are close enough in quality that availability, price per second, controllability and licence terms decide the choice, and an open-weights fallback is no longer optional for anything you ship.

The Reference Architecture: Latent Diffusion Transformers

Modern text-to-video models compress video into a small spacetime latent with a 3D variational autoencoder, then train a diffusion transformer to remove noise from that latent while attending to a text embedding. Decoding the cleaned latent yields frames. Audio, upscaling and editing are added as extra heads or stages.

Figure 1: The common latent video diffusion pipeline. Training encodes real video to latents; inference starts from noise and decodes the denoised latent.

Figure 1 shows the shape shared by Kling, Veo-class systems and the open Wan models, even though each vendor discloses different amounts. Kling is documented as a diffusion transformer with a self-developed 3D VAE that compresses space and time jointly. Alibaba’s Wan2.2 is open, so its design is inspectable. Google has published far less about Veo’s internals, so claims about its architecture should be treated as inference, not fact.

Why a 3D VAE comes first

Pixels are far too numerous to denoise directly. A 5-second 1080p clip at 24 frames per second has 120 frames of 1920 by 1080 by 3 values, roughly 746 million numbers. A video VAE learns to squeeze that into a latent grid, typically reducing height and width by a fixed factor and time by a smaller one.

An illustrative example (the factors are typical of published open models, not any one vendor’s disclosed numbers): take 720p, 5 seconds, 16 frames per second, so 80 frames. A VAE with 8x spatial and 4x temporal compression yields about 21 latent frames of 90 by 160 cells. If the transformer groups 2 by 2 latent cells into one token, each latent frame is 45 by 80, or 3,600 tokens, and the full clip is about 75,600 tokens. That is a long sequence for a transformer, comparable to a book chapter, to be denoised many times.

This arithmetic explains almost everything about the industry. Self-attention cost grows with the square of sequence length, so doubling clip duration quadruples the attention work in the naive case. Higher VAE compression, windowed or factorised attention, and distilled few-step samplers are the levers vendors pull to afford longer or higher-resolution clips. It is also why 8-second clips (Veo) and 5-to-15-second clips (Kling) are normal while minute-long single-pass generation remains rare.

Diffusion transformers and the denoising loop

A diffusion transformer, or DiT, takes the noisy latent tokens plus a timestep and a conditioning signal, and predicts the noise or, in newer flow-matching variants, a velocity that moves the sample toward clean data. Text conditioning enters through cross-attention or by concatenating text tokens into the same sequence. Modern systems often use a large language-model-derived text encoder, which is why prompt following improved sharply after 2023.

The loop runs for tens of steps at inference. Each step is a full forward pass over all 75,600 tokens in the example above, so the product of steps, tokens and layers sets the GPU bill. Distillation, which trains a student to take fewer steps, is the main reason “Fast” and “Turbo” tiers exist and cost a fraction of flagship tiers.

What is added on top

Three additions define the 2026 generation. First, audio: Veo 3 introduced synchronized dialogue, sound effects and ambient sound generated with the video, and Kling’s 3.0 Turbo tier is listed with audio included in its per-second price. Second, multi-shot and reference control, where several images or clips anchor identity so a character survives a cut. Third, editing models: Kling’s Omni line takes an input video plus instructions and returns an edited clip, with editing input and output extended to 3 to 15 seconds and 4K support according to a June 2026 release summary from the hosting provider Atlas Cloud.

Open models show the idea of expert routing. Wan2.2’s A14B variants use two experts totalling 27 billion parameters with 14 billion active per step: one handles early, high-noise denoising that sets layout, the other handles late, low-noise refinement of detail. That is a neat match between model capacity and the stage of the denoising trajectory, and it keeps inference cost close to a 14B dense model.

Training data, and what is not disclosed

Every vendor trains on large mixes of licensed, public and synthetic video with captions generated by vision-language models. Alibaba reports that Wan2.2 used 65.6% more images and 83.2% more videos than Wan2.1. Kuaishou, Google and OpenAI have not, to my knowledge, published exact dataset sizes or sources for their flagship video models. Treat any specific hour-count or parameter count quoted for Kling or Veo in blog posts as unverified unless it links to a technical report.

The 2026 Landscape: Kling, Veo, Open Weights and the Sora Gap

The practical question is no longer “which model is best on a reel” but “which model can I depend on, at what price, with what control”. This section walks the main options using only facts I could verify against primary or well-sourced material, and flags the rest.

Kling: from challenger to default hosted option

Kling belongs to Kuaishou, a Beijing short-video company founded in 2011. Version 1.0 launched in beta in June 2024 inside the KuaiYing editing app, and the line has moved quickly: 1.6 in December 2024, 2.0 in April 2025, 2.1 in May 2025 (which introduced distinct quality modes) and 3.0 as a stable release on 7 February 2026. The architecture is a diffusion transformer with a 3D VAE, as above, and the company emphasises full spatiotemporal attention to capture fast motion and scene changes.

By mid-2026 the product had split into tiers. On 17 June 2026 a hosting provider reported a new Kling 3.0 Turbo, described as a speed- and cost-optimised model with audio included, supporting 720p and 1080p. The listed price was 0.8 yuan per second of output at 720p and 1 yuan per second at 1080p, roughly 11 and 14 US cents per second at the approximate conversion the provider gave. The same release upgraded the Omni editing model, which already offered 4K and up to 15-second generation since its February 2026 launch.

Two cautions apply. Several third-party guides circulate specifics such as exact Elo scores, 60 frames-per-second modes and subscription prices, but when I traced them I found they cite no leaderboard or vendor page, so I have left them out. Also, Kling’s content moderation reportedly blocks politically sensitive subjects in line with Chinese regulation, and in May 2025 fake Kling websites were used to distribute infostealer malware. Always reach Kling through its official domains (kling.ai internationally, klingai.com in China) or a reputable API reseller.

Veo: the premium tier with an aggressive low end

Google’s Veo line was announced at Google I/O in May 2024. Veo 3 (May 2025) was the step that mattered commercially because it generated dialogue, effects and ambience in the same pass. Veo 3.1 followed in October 2025. Clips are capped at eight seconds, with longer sequences stitched in Google’s Flow tool.

Google’s pricing now spans three tiers. Per The Decoder’s coverage of the 31 March 2026 launch of Veo 3.1 Lite, the list prices per second of output were as follows.

Tier	720p	1080p	4K
Veo 3.1	$0.40	$0.40	$0.60
Veo 3.1 Fast	$0.10 (from 7 April)	$0.12 (from 7 April)	$0.30 (from 7 April)
Veo 3.1 Lite	$0.05	$0.08	not offered

Lite supports 4, 6 or 8 second clips in portrait and landscape and, according to Google, costs under half of Fast at similar speed. Google does not publish a quality difference between tiers, which means you must measure it on your own prompts. Prices change often; check the Gemini API pricing page before budgeting.

Sora: a cautionary tale, not a competitor

For two years Sora set the narrative benchmark. Sora 2 shipped in late 2025 with an API whose Pro model listed $0.30 per second at 720p, $0.50 at 1792 by 1024 and $0.70 at 1080p. All of it is now gone: the app closed on 26 April 2026, the API on 24 September 2026, and OpenAI states there is no one-to-one replacement. Any pipeline built on those endpoints had a roughly six-month migration window from the March announcement.

The lesson is structural. A frontier video model costs more to serve per user-minute than a text model, usage is bursty, and a vendor with a competing priority for the same GPUs can reprioritise. Plan for that.

Open weights: Wan, LTX and the self-hosting option

Alibaba’s Wan2.2, released on 29 July 2025 under the Apache 2.0 licence, provides T2V-A14B, I2V-A14B and TI2V-5B models. The A14B models produce 5-second 720p clips at 16 fps and recommend at least 80 GB of VRAM on a single GPU, while the 5B hybrid generates 720p at 24 fps and, per Alibaba, produces a 5-second clip in under nine minutes on a single RTX 4090. Weights are on Hugging Face and ModelScope with ComfyUI and Diffusers support.

Lightricks’ LTX-2 is a 19-billion-parameter model that generates synchronised video and audio in one model, released under its own community licence rather than Apache 2.0, so check commercial terms. A newer LTX-2.3 is listed as available. I did not verify whether later Wan versions (2.5 and beyond) have open weights, so I do not treat them as self-hostable here.

Open models trail the best hosted ones on duration, resolution and polish, but they offer fixed behaviour, no deprecation risk and fine-tuning. For regulated or brand-sensitive work that predictability is often worth more than a few points of visual quality.

Figure 2: A selection flow for AI video generation. Audio needs and hosting constraints narrow the field before price does.

Figure 2 encodes the order in which I would make the decision. Audio-in-the-same-pass is a hard requirement for many dialogue use cases and only some models offer it; hosting is a legal question before a technical one; per-second price only decides among the survivors.

Deeper Analysis: Cost, Control and Measuring Quality

What a clip really costs

List prices are per second of output, but you pay for attempts, not deliveries. If one in four generations is usable, the effective cost of a usable second is four times the list rate. Using the verified list prices above, an 8-second 1080p deliverable costs, per attempt, $3.20 on Veo 3.1, $0.96 on Veo 3.1 Fast and $0.64 on Veo 3.1 Lite. At an illustrative 25 percent keep rate, those become $12.80, $3.84 and $2.56 per kept clip. Kling 3.0 Turbo at 1 yuan per second gives 8 yuan per attempt at 1080p, about $1.12 at the provider’s approximate rate, or roughly $4.50 per kept clip on the same assumption.

The keep rate is the number you cannot read off a price page. It depends on prompt discipline, how many references you supply and how forgiving your brief is. A common pattern is draft on the cheap tier at 720p, lock the composition, then re-render only the winning prompt on the premium tier. That makes the effective cost closer to the cheap-tier rate times the number of drafts plus one premium render.

Workflow	Attempts per kept clip (illustrative)	Rough cost per kept 8s 1080p clip
All premium (Veo 3.1)	4	$12.80
All Fast	4	$3.84
Draft on Lite 720p, finish on Veo 3.1	3 cheap drafts plus 1 final	$1.20 plus $3.20 = $4.40

The numbers in the last two rows are arithmetic on list prices with assumed attempt counts, not measurements. Run your own pilot.

Controlling the output

Control is where hosted tools differentiate. Image-to-video pins the first frame, and many systems add last-frame or multi-reference conditioning so a character’s face, wardrobe or product stays stable. Motion brushes and camera presets, which Kling and several rivals offer, trade prompt ambiguity for explicit geometry. Editing models like Omni accept a source video and an instruction, which solves a different problem: changing a shot without regenerating it.

For production, think in terms of what you fix and what you leave free. Fix identity with references, fix framing with a start image, fix duration with the model’s native clip length, and leave motion and lighting to the model. Chaining 8-second clips for a longer scene works, but drift accumulates at every join, so end each clip on a neutral pose and re-anchor with the last frame as the next first frame.

Request lifecycle in a hosted API

Most commercial APIs are asynchronous because a clip takes tens of seconds to minutes. Understanding the lifecycle prevents bugs in your integration.

Figure 3: Asynchronous job lifecycle for a hosted text-to-video API, from submission to signed-URL download.

You submit parameters and receive a job id. The provider queues a GPU job, applies safety filtering to the prompt and sometimes to the output frames, writes the result to object storage with a visible or invisible watermark, and returns a signed URL that expires. Three practical rules follow. Poll with backoff or use webhooks where available. Download and re-host results promptly, because signed URLs expire. Treat a “blocked” response as a normal outcome and surface it to the user instead of retrying blindly.

Measuring quality without fooling yourself

There are two broad families of evaluation. Automated suites such as VBench score generated videos along dimensions with models and heuristics. VBench-2.0, described in an arXiv paper by a team including Yu Qiao and Ziwei Liu, targets what it calls intrinsic faithfulness across five dimensions: human fidelity, controllability, creativity, physics and commonsense. It moves beyond the earlier VBench focus on per-frame aesthetics, temporal consistency and basic prompt adherence, using generalist vision-language models, specialist anomaly detectors and human annotation to check alignment with human judgment.

The second family is crowdsourced preference arenas, where raters pick the better of two clips. They correlate with how people feel about a clip but are sensitive to prompt mix, style and the rater pool, and rankings reshuffle whenever a new model lands. I deliberately quote no leaderboard scores in this post, because the ones I found were from aggregator pages without a stable methodology link and would be stale within weeks.

The right internal approach is neither. Build a fixed set of 30 to 50 prompts that represent your real use, covering faces, hands, text on screen, fast motion and product shots. Score each output on a short rubric by two reviewers, record the keep rate and cost, and re-run the set whenever a vendor announces a new version. That is a benchmark that reflects your risk.

Trade-offs, Gotchas, and What Goes Wrong

Physics and identity still break. Models learn the statistics of motion, not mechanics. Typical failures are hands that gain or lose fingers, liquids that ignore containers, objects that pass through each other, and background characters whose faces melt when they turn. VBench-2.0 exists precisely because earlier suites rewarded smooth frames while missing these errors. Short, simple actions with one subject succeed far more often than crowded multi-step choreography.

Text and logos are unreliable. Rendered text in video frequently warps between frames. If a shot needs a legible brand mark or caption, composite it afterwards in an editor, or supply the mark through a reference image and verify every frame.

Duration is a design constraint. An eight-second ceiling on Veo, and clips of a few to fifteen seconds elsewhere, push you toward editing rather than single takes. Continuity across stitched clips is your responsibility. Budget time for colour matching and audio continuity, since each generation may differ in grain, exposure and room tone.

Prices and models are volatile. Veo Fast pricing changed within a week of the Lite launch, Kling ships a tier every few months, and Sora’s API disappeared altogether. Never hard-code a vendor into a business process. Wrap each provider behind a thin interface: prompt, reference assets, duration and aspect ratio in; a file and metadata out. Then swapping vendors is a configuration change and an evaluation run, not a rewrite.

Moderation differs by jurisdiction. Hosted tools apply different content policies, and some align with local regulation, which can block prompts that are lawful elsewhere. Test your actual subject matter early. Self-hosted open weights move that responsibility to you, including preventing misuse such as non-consensual likenesses.

Licensing is not uniform. Wan2.2 is Apache 2.0, which permits commercial use with the usual notice requirements. LTX-2 uses a custom community licence. Hosted outputs are governed by each vendor’s terms, which vary on commercial rights, indemnity and training use of your uploads. Read them with counsel before shipping client work.

Security hygiene. Fake Kling sites distributed malware in 2025. Install nothing from a “Kling desktop downloader” and use only official domains or established API gateways, with scoped keys that you rotate.

Provenance, Watermarking and Disclosure

As realism improves, provenance becomes part of the product. Two mechanisms are in use and they solve different problems. Invisible watermarks, such as Google’s SynthID, are embedded in the pixels or audio and are designed to survive some editing; they let a vendor detect its own output but do not carry rich metadata. Content credentials, defined by the Coalition for Content Provenance and Authenticity (C2PA), attach a signed manifest describing who made an asset and which tools touched it, but metadata is easily stripped by re-encoding or platform upload.

Neither is a complete defence, and neither proves a clip is real if absent. The robust approach combines them: keep the signed manifest in your master files, rely on watermark detection where a vendor offers it, and apply a visible “AI-generated” disclosure where platform rules or law require. Several platforms and some jurisdictions now require labelling synthetic media, though rules differ and change, so check the current policy of every destination. I did not verify the current C2PA specification version for this post; consult the C2PA specifications page for the latest.

Provenance also protects you. If a client questions whether a shot was synthetic, a retained manifest, prompt log and seed history is evidence. Store prompts, model versions, parameters and timestamps with each asset as a matter of routine.

Figure 4: A review pipeline for AI-generated clips, ending in provenance attachment and disclosure.

Figure 4 shows a pipeline I would run before any clip leaves the building. Cheap automated checks catch flicker and identity drift first. A vision-language judge scores prompt adherence, which is imperfect but useful for triage across many candidates. A human then checks physics, hands, brand safety and legal flags, and only approved clips receive provenance metadata and a disclosure label.

Where the Field Is Heading

Three trends look durable given what shipped in the past year. Audio-native generation is now the baseline for premium tiers, which pushes quality evaluation to include lip-sync and sound design as well as pixels. Editing models that take an existing video as input are growing faster than pure text-to-video because they fit real workflows, where the footage already exists. And price per second is falling at the low end, with Veo Lite at $0.05 per second at 720p and Kling Turbo near $0.11, which makes drafting nearly free and shifts the cost burden to human review time.

I would treat longer single-pass generation as the open question. Unless attention costs fall further, expect continuity tools, such as shot-level planning and reference chaining, to carry minute-long narratives rather than one giant clip. If OpenAI’s remark that Sora continues as world-model research is a signal, the next jump may come from models that simulate scenes with explicit state rather than painting frames, but that is speculation, not an announcement.

Practical Recommendations

Start from the use case, not the leaderboard. For short social ads and concept visuals, a hosted mid tier such as Veo 3.1 Fast or Kling 3.0 Turbo gives the best cost per kept clip. For dialogue-led scenes where audio must match lips, use a tier that generates audio natively and still plan a human dub for anything legal or safety-critical. For anything with fixed look and strict compliance, evaluate Wan2.2 or LTX-2 on your own hardware and accept lower polish for predictability.

Run a two-week pilot with the 30-to-50-prompt set described earlier and track three numbers per vendor: keep rate, cost per kept second and reviewer minutes per kept clip. The third is usually the real bill. Re-run the pilot on every major model release, and keep a second vendor warm.

• Wrap every provider behind one internal interface and store the vendor name, model version and seed with each output.
• Pilot on a fixed prompt set; record keep rate and cost per kept second, not list price.
• Draft cheap (Lite, Fast or Turbo at 720p), finish on a premium tier only for approved compositions.
• Keep an open-weights fallback validated on your own hardware.
• Re-host outputs immediately; signed URLs expire.
• Retain prompts, parameters and C2PA manifests for every published clip.
• Add a visible AI disclosure wherever the platform or law requires one.
• Review vendor terms for commercial rights and upload training clauses before client delivery.
• Never install software from unofficial Kling domains.

Frequently Asked Questions

What is AI video generation and how does it work?

AI video generation produces short clips from text, images or existing video using generative models. Most current systems compress video into a latent representation with a 3D autoencoder, then run a diffusion transformer that removes noise step by step while attending to the prompt. A decoder turns the final latent back into frames, and some systems also generate synchronised audio or upscale to higher resolution.

Is Sora still available in 2026?

No. OpenAI shut the Sora web and app experience on 26 April 2026 and the Sora API on 24 September 2026. OpenAI’s model documentation states that the Sora 2 models and Videos API are no longer available and that there is no one-to-one replacement. Sora is reported to continue internally as world-model research. If you relied on it, migrate to another hosted model or to open weights.

How much does AI video generation cost per second?

It varies widely by tier. Google lists Veo 3.1 at $0.40 per second at 720p and 1080p, Veo 3.1 Fast at $0.10 to $0.12, and Lite at $0.05 to $0.08 for 720p and 1080p. A hosting provider listed Kling 3.0 Turbo at 0.8 to 1 yuan per second, roughly 11 to 14 US cents. Real cost depends on how many attempts you need per usable clip.

Which is better, Kling or Veo?

Neither wins universally. Veo 3.1 offers a premium tier with strong audio and a cheap Lite tier, but caps clips at eight seconds. Kling 3.0 offers longer clips, multi-shot and editing models through its Omni line, and competitive per-second pricing. Independent leaderboards reshuffle constantly, so test both on your own prompt set and compare keep rate, cost and review time.

Can I run AI video generation models locally?

Yes, with open weights. Alibaba’s Wan2.2 is released under Apache 2.0; its 5-billion-parameter hybrid can generate a 5-second 720p clip on a single RTX 4090 in under nine minutes according to Alibaba, while the 14B variants recommend 80 GB of VRAM. Lightricks’ LTX-2 is a 19B model under a community licence. Expect lower polish and shorter clips than top hosted models.

How do I tell whether a video is AI-generated?

No single method is reliable. Look for C2PA content credentials, which carry a signed manifest but can be stripped, and vendor watermark detectors such as SynthID where offered. Visual cues like warped text, inconsistent hands or physics errors help, but improve less each year. Treat provenance metadata as evidence of origin when present, and its absence as inconclusive.

Further Reading

• Kling O1: the unified AI video model for consistency and editing on this site.
• Mixture-of-experts LLM architecture in 2026 for the routing idea behind Wan2.2’s two experts.
• Fact-check of viral AGI videos on judging synthetic media claims.
• VBench-2.0 paper on arXiv for the evaluation dimensions discussed above.
• OpenAI’s Sora 2 Pro model page for the shutdown notice and historic pricing.
• The Decoder on Veo 3.1 Lite pricing for the tier table.

By Riju — about