The second wave — Runway, Kling, Luma, Pika
Below Google's Omni / Veo 3.1 stack sits a second wave of models with documented APIs. They are not "best in class" on every axis but they are the right answer for a lot of real production jobs. This is where volume work happens, because Veo 3.1 Standard at $0.40/sec is too expensive for a 200-clip test set.
Runway Gen-4.5 + gen4_turbo — the production workflow
Runway has been in the AI-video game since Gen-1 in 2023 and the company's edge has always been workflow, not raw model quality. Two documented rates:
- gen4.5: $0.12/sec
- gen4_turbo I2V: $0.05/sec
A 5-second gen4_turbo I2V clip is $0.25 in raw API spend —
same math as Veo 3.1 Lite at $0.05/sec. Gen-4.5 is the
flagship; Turbo is the cheap I2V.
Runway also ships Aleph (a video-to-video editing model), audio sync tools, and a real timeline UI. The value is being able to generate, edit, mask, and finish in one product instead of juggling Veo 3.1 + DaVinci + ElevenLabs. For teams already inside Runway it remains the default; for new teams Omni / Veo 3.1 direct is usually cleaner per dollar but loses the workflow.
Kling 3.0 — the value heavyweight
Kuaishou released Kling 3.0 on February 5, 2026. The live family
name on the API is kling-v3. The model uses a
Multi-modal Visual Language (MVL) architecture — text, image,
audio, and video all processed in one system — and ships native
4K output with built-in multilingual audio.
Pricing:
- $0.084/sec standard mode (no video input)
- $0.168/sec Pro mode (video input or higher quality)
- Subscription tiers: Standard $6.99 / Pro $25.99 / Premier $64.99 / Ultra $127.99-180 per month
This is the price-performance leader in the heavyweight class. Kling 3.0 at $0.084/sec is roughly 20% the price of Sora 2 Pro at $0.50/sec for output that, on the right kind of prompt, is comparable. Where Kling loses: long-duration character consistency and the physics edge Sora 2 still has through 2026-09-24.
Access caveat: the global onboarding flow is clunky for non-Chinese developers; Atlas Cloud and other aggregators re-sell the API with simpler auth (and at lower prices).
Luma Ray 3.2 — image-to-video motion
Luma Labs' Dream Machine and the Ray family (current: Ray 3.2,
Agents API) focus on smooth motion from a starting image. The
image-to-video sweet spot — you compose the first frame with
gemini-3.1-flash-image via generate_content, then Luma
animates it.
Use Luma when the brief is "this image, in motion, smoothly." Avoid for character dialogue, complex physics, or anything where the camera move matters more than the subject. Check current Agents pricing before you budget; do not copy an old per-clip number out of a blog post.
Pika — aggregator, not a first-party cheap model
Pika's lane is creative effects and dialogue, sold as a subscription UI on top of other people's video models. Two surfaces still matter:
- Pikaffects: dramatic transformations applied to images or clips with single-prompt presets (melt, explode, inflate, squish, morph). Great for TikTok/Reels-shaped content.
- Pikaformance: audio-driven lip-sync. Take a single still image, give it a voice clip, get a talking head. Comparable to HeyGen for localized character content.
You are not buying a cheap first-party model. You are buying the
aggregator. Do not put pika in a router as if it were Veo 3.1
Lite.
The honest summary
If you can only remember one rule: Omni / Veo 3.1 for Google,
Kling 3.0 (kling-v3) for value at quality, Veo 3.1 Lite or
Runway gen4_turbo I2V for cheap volume, Runway gen4.5 for
editing workflow, Higgsfield when camera matters, Luma Ray 3.2
when motion-from-image matters, Pika when the effect UI is the
product, Sora 2 only through 2026-09-24 and not as the
production default.
That's not a leaderboard; it's a lane map. The lanes are durable. Lane leaders change every six months.