AI Video Generation APIs Compared (October 2026): Gemini API, Runway, Kling and Luma

Short answer: with Sora’s API gone since September 24, 2026 (OpenAI deprecations), the main official video generation APIs are Google’s Gemini API (Gemini Omni Flash and Veo 3.1), Runway’s API (Gen-4.5 plus partner models), the Kling API (Kling 3.0 family) and the Luma API (Ray3.2). Google is the most direct choice for prompt-to-clip with audio. Runway is the easiest way to reach many models through one integration. Facts below come from each vendor’s developer docs, checked on October 18, 2026.

Which AI video APIs can you use today?

These four are documented on the vendors’ own developer sites:

API	Current video models	Auth	Billing unit	Docs
Gemini API (Google)	gemini-omni-1.1-flash (generally available on the paid tier); veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview	API key (x-goog-api-key)	Veo: per second of video, by resolution; Omni Flash: output tokens, which Google converts at a fixed rate per second of 720p video	Video overview
Runway API (Runway Dev)	Gen-4.5, Aleph 2.0 and partner models: Gemini Omni Flash, Veo 3.1, Kling 3.0 and O3, Seedance, MiniMax, Wan, Grok Imagine and more	Authorization: Bearer plus X-Runway-Version header	Credits per second, bought per project in the developer portal	Runway Dev
Kling API	Kling 3.0 Turbo, 3.0, 3.0 Omni, O1, 2.6, 2.5 Turbo	Authorization: Bearer <API key>	Units per second, deducted from resource packages	Kling quick start
Luma API	Ray3.2 (video), Uni-1.1 (image)	See Luma’s docs	Credits; cost varies by model and output	Luma

If you’re migrating off OpenAI, read Sora API shutdown: how to migrate first.

What does each API do best?

Gemini API: Omni Flash or Veo 3.1?

Google tells developers to “use Gemini Omni Flash as your default model for video generation”, and to use Veo 3.1 “for specific capabilities like scene extension, last-frame control, or integration with legacy pipelines” (Gemini API).

• Omni Flash takes text, image, audio and video input together. It supports multi-turn editing through the Interactions API and first-and-last-frame interpolation. It outputs 720p by default, a 360p option, and upscaled 1080p and 4K (Omni docs).
• Veo 3.1 makes 4, 6 or 8 second clips with native audio at 720p, 1080p or 4K. 1080p, 4K and reference images require 8-second clips, and Lite has no 4K and no reference images. You can extend Veo videos by 7 seconds at a time on Veo 3.1 and Fast (not Lite), up to 148 seconds in total (Veo docs).
• Operational limits worth planning for: Veo requests take from 11 seconds to 6 minutes at peak. Generated videos are deleted from Google’s servers after 2 days, and every Veo video carries a SynthID watermark (Veo docs).
• Preview status: the three Veo 3.1 model IDs are labeled Preview (Veo docs).

Runway API: why use an aggregator?

One key, one task-polling pattern, many models. Runway’s catalog describes “the best video, image, audio, and real-time models, available through one API” (catalog). Things to know from its API primer:

• Generation is asynchronous: you get a task ID, then poll GET /v1/tasks/{id}.
• Request fields change per model. gen4.5 accepts durations from 2 to 10 seconds; gemini_omni_flash accepts 3 to 10. Runway warns: “never carry a ratio or duration from one model to another”.
• Content moderation shows up as a FAILED task, not an HTTP error, and “moderated generations cost the same as successful ones”.
• Official SDKs: @runwayml/sdk (Node) and runwayml (Python).

Kling API: what’s covered?

The Kling API covers Kling’s 3.0 family and older models, plus avatar, lip sync and audio endpoints (capability map). Kling 3.0 and 3.0 Omni generate 3–15 second clips up to 4K, and support first and last frames for image-to-video.

• Kling 4.0 is not on the API yet. Kuaishou says Kling 4.0 “will soon be made available to enterprise customers and developers through API access” (Kuaishou).
• Getting a key: sign in at kling.ai/dev with your Kling account, buy a resource package (a trial package exists for integration testing), then create a key. Kling says the key “will only be shown once” (quick start).
• Host: https://api-singapore.klingai.com.

Luma API: what’s different?

Luma’s API is built around Ray3.2. It exposes Multi-Keyframe with up to 16 keyframes, video-to-video up to 20 seconds, and Reframe. It offers Python and JavaScript/TypeScript SDKs and two tiers: Build (usage-based) and Scale (dedicated capacity with an SLA) (Luma).

How is AI video API usage billed?

Almost always per second of output, with multipliers for resolution, audio or video input. We don’t list amounts; use the official pages.

Tier	Best for	Key differences	Official pricing
Gemini API Free Tier	Text and image models	Veo 3.1 and Gemini Omni Flash are listed as “Not available” on the free tier	Gemini API pricing
Gemini API Paid Tier	Veo 3.1 or Omni Flash in production	Needs a linked billing account; Veo Lite is the cheapest Veo tier and Standard the most expensive; 4K costs more	Gemini API pricing, billing
Runway API credits	Many models through one key	Each model has its own credits-per-second rate; Model Router calls bill at the chosen model’s rate	Runway API pricing
Kling API resource packages	Kling 3.0 family	Per-second units that rise with resolution, native audio and video input	Kling API pricing
Luma API Build / Scale	Ray3.2 keyframes and video-to-video	Build is usage-based with no commitment; Scale adds dedicated capacity and an SLA	Luma pricing, Luma API info

For a sense of how token-based pricing works on other APIs, see OpenAI API pricing and DeepSeek API pricing.

Can you use API output commercially?

Most API terms let you use the output in your product, with conditions. Summary only, not legal advice.

• Gemini API: “Google won’t claim ownership over that content”. Google also says it “may generate the same or similar content for others”, and that the service is “for developers building with Google AI models for professional or business purposes” (Gemini API terms).
• Runway: it “does not restrict your commercial use of your Outputs”. Its terms also grant Runway a license to use inputs and outputs to train its models (Runway Terms).
• Kling API: “You [sic] use of the AI-generated content for commercial purposes is not restricted”, with a reminder that AI output can still infringe third-party rights (Kling API terms).
• Luma: Luma assigns ownership of output to the customer, but says you “can only use the Outputs for commercial purposes if the Outputs were produced during an active Subscription Term under Customer’s paid subscription allowing for the commercial use of those Outputs” (Luma Terms).

More detail by tool is in Can you use AI videos commercially?

How do you choose a video API?

1. List the controls you need: clip length, last frame, references, audio, editing an existing clip.
2. Check model status: preview models (Veo 3.1) can change, and newer models (Kling 4.0) may reach the app before the API.
3. Prototype on two vendors. An aggregator such as Runway lets you swap models without a second integration.
4. Plan for async jobs and retries. All of these APIs return a job or task to poll. Rate limits and moderation failures need handling; see fix AI API errors.
5. Save outputs right away. Google deletes generated videos after 2 days, for example.

For a model-level comparison rather than API details, see Veo 3.1 vs Kling vs Runway. All AI video guides are in the AI video tools hub.