Sora 2 vs Veo 3.1 vs Runway Gen-4 vs Kling: which AI video tool to use for which job in 2026

OpenAI is shutting Sora 2 down. The web app closed on April 26, 2026. The API ends on September 24, 2026. There is no announced successor. Press coverage of the wind-down points to compute costs in the range of a million dollars a day and a daily-active-user drop from around one million to under half that (The Decoder).

If you have a Sora pipeline in production, the window is now about four weeks. As of late August 2026 that is not enough time for a full evaluation, so port your highest-volume job first and treat the rest as follow-up. If you are about to start a new AI video build, Sora is off the table.

That leaves three viable shortlist tools that buyers ask us about most weeks: Veo 3.1, Runway Gen-4 / Gen-4.5, and Kling 3.0 / 3.0 Omni. The trick is that they are not interchangeable. Each one wins on a different axis. Pick the wrong one for the job and you spend the next quarter fighting the model instead of shipping work.

This is the matrix we run when a client asks us to scope an AI video build. Use cases, named production examples, real per-clip pricing, failure modes, and a migration plan if you are still on Sora. We are going to be honest about where each tool is weak and where the newer cohort outside the four is already pulling ahead.

The four-way wins-on framework, at a glance

If you only read one table on this page, read this one.

The usual one-liner from forum threads (audio = Veo, physics = Sora, volume = Kling, control = Runway) is still right. The 2026 refinement is that Kling's 3.0 Omni release in February pulled it into Veo's audio lane on the leaderboards, and native 4K and multi-shot dialogue split into separate categories that did not exist a year ago.

Twelve jobs, one model recommendation each

The matrix below maps real jobs that come up in our scoping calls to a primary tool. Each row maps a common job to the tool we would reach for first.

The four sections that follow go deeper on each tool. We are leading with Veo and Runway because those are the two we recommend most often in 2026 buyer scopes.

Veo 3.1: cleanest audio and the only native 4K of the four

Google's Veo 3.1 (and its faster sibling, Veo 3.1 Fast) gained native 4K on January 13, 2026. The audio stack runs at 48 kHz with lip-sync latency around 120 ms, and prompt adherence on dialogue is consistently the best of the four in blind testing.

What you actually get per clip

• 8-second native clips
• Up to roughly 140 seconds via chained extensions, but only at 720p
• 4K 3840 x 2160 native, 16:9 and 9:16
• Audio at 48 kHz with sub-150 ms lip-sync
• Up to three reference images for character or style anchoring

Pricing reality

Veo runs on credits inside Google AI plans. The catch is that credits do not roll over. The real cost per clip is the monthly plan price divided by the number of clips you actually make, not the cap.

• Flow free tier: 50 generations per day
• Google AI Pro: $19.99 per month, 1,000 credits, around 90 Fast videos
• Google AI Ultra: $200 per month, 25,000 credits (or the $100/month Ultra tier, 10,000 credits), around 2,500 Fast videos

On the Ultra plan amortised at full use, a 30-second 1080p Veo clip is roughly $2.40. At half use, it is $4.80. Build your forecast on what the team will actually generate, not the headline cap (Google AI plans, Flow credits explainer).

Named cases

• Pit Viper x Dept ran a Veo-3 pipeline for a sunglasses brand campaign that hit a 90% cost reduction versus their traditional shoot (Think with Google case study).
• Hector the Hare, a short narrative film built largely on Veo, showed what the model can do on a single-character story with controlled lighting.
• Coign Credit Card ran what is widely cited as the first fully AI-generated television commercial in the US, using Veo for the bulk of the production.

Where Veo struggles in production

The most common complaint we hear is the look of skin, hair, and crowds in mid-shots. It is not bad, but it is recognisable. Physics is behind what Sora used to deliver. Vertex AI rate caps (10 requests per minute on default tiers) become a real constraint on production runs. Non-English dialogue is noticeably weaker than English. Several teams have flagged a content-safety regression in early 2026 where prompts that worked in December started getting refused in February.

When we recommend Veo

Broadcast and CTV commercials that need a 4K master, single-shot or short-form social with dialogue, and any work where audio quality is the deciding factor.

Runway Gen-4 / Gen-4.5: control, character consistency, and the only studio distribution

Runway is the only one of the four with named major-studio and creative-suite partnerships. Lionsgate trained a custom catalog model on a 20,000-title library. AMC Networks signed the first cable-TV-first AI deal in June 2025. Adobe added Runway to Premiere and After Effects in December 2025 (Adobe news, Hollywood Reporter on AMC, Variety on Lionsgate).

That changes the buying conversation. It stops being about which clip looks best and starts being about which vendor your legal and post-production teams can already approve.

Capability surface

• 5- to 16-second clips at 720p or 1080p
• Motion Brush, Camera Control, Director Mode
• Act One for performance capture from a reference video
• Strong reference-anchored character consistency, the lowest identity drift of the four
• Limited native audio (you usually pair with a separate audio stack)

Pricing

• Free 125 credits one-time
• Standard $15 per month (625 credits)
• Pro $35 per month (2,250 credits)
• Max $95 per month, or $76 per month billed annually (9,500 credits)
• Enterprise custom
• Gen-4.5 consumes 25 credits per second of generated video

At the Pro tier, a 30-second clip lands around $11.67. Enterprise pricing is by quote and folds in indemnification (Runway pricing, API pricing).

Where Runway is weaker

Photorealism still trails Veo on the live-action side. Audio is the biggest gap, which is why brand films built on Runway often end up as a two-tool pipeline (Runway for picture, Veo or a separate text-to-speech stack for sound). And the Lionsgate partnership, much-celebrated at launch, ran into real production constraints in its second year (3DVF coverage). The lesson is that a custom-trained studio model is a long-cycle programme, not a plug-in.

When we recommend Runway

Character-driven narrative work, brand films, anything that has to clear post-production approval inside a Premiere or After Effects-led workflow, and any client whose legal team is more comfortable with a US-headquartered vendor with public IP licensing language. We default to Runway for storyboard migrations off Sora.

Kling 3.0 / 3.0 Omni: multi-shot dialogue and the cheapest per-clip in the matrix

Kuaishou released Kling 3.0 and the Omni variant on February 4, 2026, with the API following the next day. The headline change was that Kling stopped being just the cheap-volume option. The Omni release added native audio, five-language lip-sync (English, Chinese, Japanese, Korean, Spanish) with accent control, and a six-shot storyboard with a synced audio timeline. Kling is now the only model of the four with native multi-shot dialogue (Kling 3.0 release, Kuaishou IR).

Capability surface

• Native audio with five-language lip-sync and accent control
• Six-shot storyboard with a synced audio timeline
• Up to 4K
• Reference-video character and voice extraction

Pricing

• Free 66 generations per day
• Standard $10 per month (660 credits)
• Pro $37 per month (3,000 credits)
• Premier $92 per month (8,000 credits)
• Ultra $180 per month (26,000 credits), no annual option as of mid-2026
• First-time subscribers get a discounted opening month (Standard $6.99, Pro $25.99, Premier $64.99, Ultra $127.99), which is not the ongoing rate
• API: $0.084 per second standard, $0.168 per second pro

At the Pro API tier, a 30-second clip is about $5.04. At standard, it is $2.52. That is the cheapest per-clip price in the four-tool shortlist by a wide margin (Felloai pricing breakdown, fal.ai Kling 3).

Named cases

• Atlas Cloud ran a case study on automating high-quality video marketing at scale on Kling 3.0 (case study).
• Picsart integrated Kling 3.0 Omni for creator-facing video generation (Picsart announcement).

Where Kling is weaker for studio buyers

The big one is procurement. Kuaishou is a Chinese parent company with infrastructure in China. For a major-studio buyer with TPN or MPA-Plus requirements, that is a real review hurdle. Kling's output also tends to be more stylised than Veo's on the same prompt, and character identity tends to drift past 12 seconds on longer clips.

When we recommend Kling

Multi-language UGC and creator work, anything dialogue-heavy that has to ship in five languages, programmatic volume at the lowest per-clip rate, and any client where the procurement bar does not include China-infrastructure restrictions.

Sora 2: how to think about it for the next four months

Sora 2 is a migration topic, not a buying topic. We are including it here because it is in the title and because if you have a pipeline running on it today, you have a real port to plan.

The shutdown timeline

• April 26, 2026: web and app discontinued (OpenAI Help Center)
• September 24, 2026: API ends (OpenAI deprecations)
• No announced successor on the public deprecations page as of May 2026

Capability while it lasts

• 4 to 25 second clips at up to 1024 x 1792
• Sora 2 standard 720p: $0.10 per second
• Sora 2 Pro 720p: $0.30 per second
• Sora 2 Pro 1024p: $0.50 per second
• A 30-second 1024p clip is around $15
• Physics realism is still the leader of the four on blind tests, especially on cloth, fluid, and collision behaviour

The proof point that is itself a proof point

In 2026 press coverage we cannot find a single named brand campaign on Sora 2. CAA flagged Sora as a "significant risk to clients" in early 2025 and has not changed that position publicly (Hollywood Reporter). The absence of named brand cases is the case. Buyers stayed away.

What to do if you are on Sora today

Port. See the migration table below. Plan a four- to six-week porting window if you are still running production on Sora at the start of Q3 2026.

Pricing reality, side by side

Real-world prices change weekly. These are the May 2026 numbers normalised to a 30-second 1080p clip where the model supports it.

Two things worth flagging. First, Veo credits do not roll over. Anything you do not use in the month is gone, so the realistic cost per clip is plan price divided by clips actually shipped, not the cap. Second, every model bills the way that flatters its own pricing page. Runway counts in credits, Kling in API seconds, Veo in plan-month buckets. If you are buying for a programme, normalise to the unit your team actually ships in (typically per-minute of final-cut video).

Failure modes by axis

No model in the four is failure-free. The honest breakdown by axis, from our scoping work and the public consistency literature, looks like this.

The two universal failure modes (text inside frames, multi-character interaction) are still where AI video as a category is weakest. Plan around them. Render text in post, and use single-character scenes or carefully blocked two-handers where you can.

Sora migration paths

If you are running a Sora pipeline today, here is the porting map we use with clients. For the full step-by-step version, see our dedicated guide on what to switch to after the Sora shutdown.

The hard deadline is September 24, 2026. If you have a production programme on Sora today and you have not started porting by July, you are going to feel it. We are scoping migrations like this inside our AI Content Production line.

What this piece is NOT covering

This piece is about the four generative video models buyers ask about by name. Three adjacent buyer journeys come up in the same conversations but need different shortlists.

• Deterministic, programmatic, brand-locked video at 1,000+ variant scale. That is Hyperframes, HeyGen's open-source HTML-to-video renderer (Apache 2.0, April 2026). It is not a generative model. It renders programmatic video from HTML and is the right answer when your job is personalisation at scale, not a single creative spot (Hyperframes GitHub).
• Real-time live AI dubbing and broadcast. Sub-second latency is a different stack from any of the four in this piece. We have a separate write-up on the real-time playbook in AI dubbing and voice cloning, and a Wave-1 piece on the MLS / FanCode / NASCAR sub-second models is coming.
• Self-hosted or on-prem open-weight video. That is Wan 2.5 / 2.6 from Alibaba. Open weights, native audio, self-hostable. If your buyer is a CISO who has blocked vendor APIs, that is the right shortlist. See our private LLM hosting write-up for the broader pattern.

What is beyond the four (and why we still recommend the four most of the time)

As of May 2026, the top of the Artificial Analysis text-to-video-with-audio arena is no longer the four models in our title. The current leaderboard reads:

• Dreamina Seedance 2.0 (ByteDance), Elo 1222
• HappyHorse-1.0, Elo 1214
• Kling 3.0 Omni, Elo 1105
• Veo 3.1, Elo 1102

From around 420,000 votes across 39 models (Artificial Analysis arena).

Why are we still recommending Veo, Runway, and Kling for most paid work?

Because buyer-side procurement does not move at leaderboard speed. Veo has Google Cloud, billing, support, and indemnification. Runway has Adobe, Lionsgate, and AMC inside their legal-approved vendor lists. Kling has the only multi-shot dialogue timeline at scale. Seedance 2.0 is real and rising, but most enterprise buyers do not yet have it on their approved list. We expect that to shift in the next two quarters. We will refresh this piece when it does.

The honest line for a buyer is: if your procurement team will sign with Dreamina or the smaller arena leaders, give Seedance 2.0 a serious look on physics-heavy and short-form work. If they will not, the four in this piece are still the right shortlist for most scopes.

A one-page decision tree

This is the version we hand to clients on a single page.

1. Is the deliverable a 4K broadcast or CTV master? Veo 3.1.
2. Is the deliverable a multi-shot brand film that has to clear post-production approval? Runway Gen-4.5.
3. Does the work need dialogue in two or more languages with lip-sync? Kling 3.0 Omni.
4. Is it under 12 seconds with character consistency as the main constraint? Runway Gen-4.5 with Act One.
5. Is it physics-led (cloth, fluid, collisions) and you can ship before September 24? Sora 2 still wins, but plan the port now. After September, Hailuo 2.3.
6. Is the job 1,000+ personalised variants? This is not a generative-model question. Use Hyperframes (deterministic) or a personalisation engine on top of one of the four.
7. Is the constraint that the vendor API cannot leave your network? Wan 2.6 (open weights) or a Runway Enterprise on-prem conversation.
8. Is the deliverable real-time, live, sub-second? None of the four. Different stack. See our AI dubbing and voice cloning write-up.

We scope these calls on the back of this tree two or three times a week. If the answer is not obvious, scope a 30-minute call with us and we will run the decision live: book a strategy call.

Keep reading

For the full 2026 video-generation field beyond the four in this piece, see our AI video generation 2026 overview. For the real-time and dubbing stack, see AI dubbing and voice cloning. For the surrounding legal and compliance work that buyers now want with every AI video brief, see our EU AI Act guide for SaaS founders and our Enterprise AI service for running these pipelines on private infrastructure that keeps your footage and prompts under your own control. If you are working on a longer-form narrative or a multi-language creator pipeline, AI Content Production is the right service line to scope against.