Sora vs Veo vs Kling vs Runway: What AI Video Really Costs in 2026

This comparison needs a different framing than it would have six months ago. OpenAI's Sora web and mobile apps shut down on April 26, 2026. The Sora API — the part relevant to anyone building on it programmatically — is scheduled to shut down on September 24, 2026. As of this writing (September 15, 2026), that's nine days away. OpenAI has given no indication of a Sora successor or timeline for one. For any planning purpose beyond the next nine days, Sora is not a viable option, and this guide treats it that way: covered for historical and migration context, not as a live recommendation.

That leaves three real contenders for anyone choosing a video API today: Google Veo 3.1, Kling, and Runway. The rest of this guide focuses on what each actually costs — not the generation price alone, but the cost of a finished, usable clip after the retries that raw generation pricing never accounts for.

Sora: the shutdown timeline, briefly

Reported drivers include compute cost (widely cited around $1M/day to run) and a usage decline after an initial peak near 1 million users. OpenAI's stated redirection is toward ChatGPT, Codex, and enterprise products. If you have an active Sora API integration, the honest planning assumption is that it needs a replacement in the next nine days, not a comparison for later.

Google Veo 3.1: current API pricing

Veo 3.1 generates native synchronized audio — dialogue, ambient sound, effects — in the same pass as the video, which matters for the cost comparison: a model that requires a separate audio-generation step elsewhere has a real cost this per-second figure doesn't show. Veo 3.1 also supports scene extension (chaining generated clips for continuous narratives beyond a single generation's length) and start/end-frame-guided generation, both relevant to multi-clip project costs below.

Access note: Veo 3.1 API pricing is available through Vertex AI (enterprise-grade, pay-per-second) and, at preview status with more limited access, through the Gemini API directly. Consumer access also exists bundled into Google AI Pro ($19.99/mo, roughly 3 Veo 3 Fast videos/day via the Gemini app) and AI Ultra ($249.99/mo, substantially higher Flow-editor allowance) — a meaningfully different cost structure from the pay-per-second API and not a fair substitute for programmatic use.

Kling: the pricing-volatility case

Kling's API pricing is, on the evidence available, the least stable of the three current options. Multiple independent trackers show the published range widening substantially over 2026: from roughly $0.08–$0.17/second in April 2026, to $0.08–$0.42/second by May, to $0.06–$0.42/second by August — a top-tier increase of over 150% within four months, alongside new tiers being added (Standard, Pro, 4K, Resource Packages, Custom) rather than a single rate being revised. Independent monitoring has also flagged Kling's pricing as fluctuating by as much as 40–60% "regularly" across model versions, according to community reports aggregated by third-party trackers.

This is an area of genuine uncertainty, stated plainly rather than resolved with a single confident number: at time of writing, Kling's per-second rate for a comparable 720p-with-audio configuration appears to sit somewhere in the $0.11–$0.17/second range based on the most recent snapshots available, but given the frequency and size of recent changes, this is the one figure in this guide most likely to be outdated by the time you read it. Check Kling's official pricing page directly before budgeting a project against any third-party figure, including this one.

Runway: subscription vs. API, two different cost structures

Runway is unusual among the three in offering a genuinely separate consumer subscription and developer API, with separate credit balances that don't share allowance.

Consumer subscription (web app):

Developer API (Runway Dev): pure pay-as-you-go at $0.01/credit, minimum $10 purchase. Gen-4 Turbo costs 5 credits/second ($0.05/sec); Gen-4.5 costs 12 credits/second ($0.12/sec); third-party models routed through Runway (including Veo 3.1) cost 20–40 credits/second ($0.20–$0.40/sec) depending on audio. HDR or high-resolution output adds a further 20–40 credits/second on top of the base rate.

FACT worth flagging: Runway's API and web-app credit balances are entirely separate. Paying for a Standard, Pro, or Max subscription does not add credits to an API account, and API purchases don't add credits to a web-app workspace. A team using both surfaces is effectively running two separate Runway budgets, not one.

The real comparison: cost per usable second, not per generated second

Every price above describes the cost of a generation, not the cost of a generation you actually keep. A generation that gets rejected — wrong motion, an artifact, a prompt that didn't land as intended — still cost money to produce. Effective cost per accepted second is the raw rate multiplied by how many attempts it typically takes to get one you'll use.

CALCULATION — Runway's own published example, from its official credit documentation: 10 clips, 5 seconds each, with an average of 5 generation attempts per accepted clip, on Gen-4.5 at 12 credits/second:

10 clips × 5 attempts × 5 seconds × 12 credits = 3,000 credits

At Runway's $0.01/credit API rate, that's $30 in generation spend to produce 10 accepted 5-second clips (50 seconds of finished footage) — a 5x gap between the "ideal" cost (10 clips × 5 sec × 12 credits × $0.01 = $6.00 at one attempt each) and the realistic cost once regeneration is accounted for.

Applying the same regeneration-multiplier logic to Veo 3.1 Standard ($0.40/sec, with audio) for the same 50 seconds of finished footage:

• Ideal (1 attempt/clip): 50 × $0.40 = $20.00
• Realistic (5 attempts/clip, matching Runway's own published assumption): 50 × 5 × $0.40 = $100.00

And for Kling at an indicative $0.13/second:

• Ideal: 50 × $0.13 = $6.50
• Realistic (5x): 50 × 5 × $0.13 = $32.50

The ranking by raw per-second price and the ranking by realistic finished-footage cost can diverge once each model's actual regeneration behavior is factored in — and that behavior is exactly the variable no vendor publishes and every buyer has to measure for themselves on a representative sample before committing a production budget.

Comparison table: raw price vs. realistic 50-second project cost

ASSUMPTION flagged clearly: the 5x regeneration multiplier applied uniformly across all five rows above is Runway's own published assumption for its specific worked example, used here for a consistent, apples-to-apples comparison — it is not a verified, model-specific regeneration rate for Veo 3.1 or Kling, and actual regeneration rates likely differ by model, prompt complexity, and use case. Treat this table as illustrating the mechanism (raw price and realistic price diverge, sometimes by a lot) rather than a precise prediction of what any specific model will cost you.

The cost of migrating off Sora

For teams with an active Sora integration, the shutdown itself carries a cost beyond picking a replacement API: re-engineering whatever pipeline called Sora's endpoints to match a different provider's request format, parameters, and output handling, plus re-testing prompts that were tuned for Sora's specific output characteristics against a new model that will behave differently even on an identical prompt. That's real engineering time, not just a rate-card comparison — budget for a prompt-retuning pass on whichever replacement you choose, since a prompt library built around one model's quirks rarely transfers cleanly to another's.

Scaling to a full project: 30 seconds and 60 seconds of finished footage

The 50-second example above illustrates the mechanism; here's the same math applied to two common deliverable lengths, using the same illustrative 5x regeneration assumption for consistency.

30-second finished sequence:

60-second finished video:

CALCULATION check (60-second Veo 3.1 Standard row): 60 sec × $0.40/sec = $24.00 ideal; × 5 attempts = $120.00 realistic. The pattern holds across every length and provider in this comparison: the gap between "what the pricing page says" and "what the project actually costs" scales linearly with the regeneration rate, which is exactly why measuring your own rate on a test batch before committing to a full production run matters more than comparing headline per-second prices across vendors.

A lower regeneration rate changes these numbers substantially — at a more optimistic 2x rate (still not "ideal," but closer to what a well-tuned prompt and workflow might achieve), the 60-second Veo 3.1 Standard realistic cost drops from $120.00 to $48.00. The vendor comparison doesn't change; the absolute stakes of getting your own regeneration rate right do.

Who should choose which provider

Runway fits teams that want the lowest raw per-second entry cost via the API (Gen-4 Turbo at $0.05/sec) and don't mind managing a separate credit balance from any consumer-app usage, or teams that want an editing-integrated workspace via the subscription tiers.

Veo 3.1 fits projects where native synchronized audio in the same generation pass is valuable enough to justify the highest per-second rate of the three — for anything requiring dialogue or precise sound design, the alternative (generating video and audio separately, then syncing them) has its own real cost that Veo's bundled pricing avoids.

Kling fits budget-conscious projects willing to accept some pricing unpredictability and to re-verify current rates immediately before committing to a production run, given the volatility documented above.

FAQ

Can I still use Sora right now? The API is live until September 24, 2026 — nine days from this article's publication date. After that, it's fully discontinued with no announced successor. Do not build anything new on it.

What's the best Sora replacement? Based on feature overlap (native audio, high resolution, prompt fidelity), Veo 3.1 is the closest match cited by multiple migration guides. Kling is commonly cited as the cheaper alternative. Runway suits editing-centric workflows better than either.

Why is Kling's pricing described as uncertain in this guide? Because independent trackers show its published rate range changing significantly and repeatedly across 2026 — a 150%+ top-tier increase within four months, per available data — making any single figure quoted today a snapshot rather than a stable reference point.

Does Runway's subscription or API work out cheaper? It depends entirely on volume and whether you need the web app's editing tools. The API's cheapest tier (Gen-4 Turbo, $0.05/sec) beats every subscription tier on a pure per-second basis; the subscription tiers bundle an editing workspace the API alone doesn't include.

Is the "5x regeneration" multiplier realistic for my project? It's Runway's own published assumption for one specific worked example, not a universal constant. Treat it as illustrative of how much raw and realistic costs can diverge, and measure your own regeneration rate on a representative sample before budgeting a real project.

The NotCheapAI verdict

With Sora's API days from shutdown, the real 2026 video-API comparison is a three-way race between Veo 3.1, Kling, and Runway — and the raw per-second price tells you less than it looks like it does. Runway's API undercuts the other two on paper; Veo 3.1's bundled native audio can offset its higher rate for audio-heavy work; Kling's price is the cheapest of the three on some configurations and the least stable across all of them. Whichever you choose, budget from a realistic regeneration rate you've measured yourself, not the number on the pricing page.