Video is no longer a research demo

You can finish this chapter in the browser without buying a second of video. Veo (Google) and the other names below are clip-making tools — type a description, get a few seconds back. The price sheet is so you can compare lanes, not a bill you owe to pass.

Image generation went from "uncanny" to "production-grade" in about two years. Video has taken roughly the same trajectory, but compressed into eighteen months: when Sora 1 leaked in early 2024 it was a spectacle. Sora 2 shipped as an API in October 2025. Veo 3.1 Lite landed March 2026 at $0.05/sec. Google's default video model is Gemini Omni 1.1 Flash (gemini-omni-1.1-flash). Kling 3.0 (February 2026, native 4K with audio) is still in the working set. There are at least eight distinct video models with real production users, each occupying a different lane.

Two dates sit on that lineup. Sora 2 / Videos API is removed 2026-09-24. gemini-omni-flash-preview is removed 2026-09-30. Until those dates the calls still run. After that, those IDs are dead. Veo 3.0 and Veo 2.0 IDs (veo-3.0-*, veo-2.0-*) already shut down on the Gemini API on 2026-06-30. The live Veo family is 3.1.

The catch: video is also the most expensive generative modality by an order of magnitude. A 60-second promo on Sora 2 Pro at 1024p costs $30 in raw API spend. That number is 60 seconds × $0.50/sec, not one request: Sora jobs land in the 16-20 second range, so a minute is three extends, and moderation failures still get billed. The same minute on Veo 3.1 Standard with audio runs $24. Veo 3.1 Lite at 720p with audio is $3. Omni is ~$6. None of this includes the "generate ten, pick one" reality. The "$3-5/finished-minute including retakes" headline only holds if you are on Lite and the retake rate is about 1.

The four questions that decide which model

1. Do you need native audio? The Veo 3.1 family — Lite, Fast, and Standard — generates speech, ambience, and music in the same pass as the picture. Omni generates an audio track by default; you prompt the sound. Sora 2 generates audio through 2026-09-24. Native audio is not a Standard-only premium. Everyone else makes you composite audio in post.
2. Do you need camera control? Higgsfield's specialty — virtual lenses, focal lengths, stacked moves, character lock-in. Most base models can't do a coherent dolly-in or rack focus. Higgsfield wraps the base models with that vocabulary. You can skip it: named shot language in the prompt is the other path.
3. Is the budget per-minute viable? Cheap tier (Veo 3.1 Lite at $0.05/sec, Runway gen4_turbo I2V at $0.05/sec): $3/min raw at 720p+audio on Lite. Omni ~$6/min at 720p. Veo 3.1 Standard: $24/min. Different products entirely.
4. Text-to-video, image-to-video, or video-to-video? Image-to-video is the sweet spot for product work — you author the first frame with gemini-3.1-flash-image via generate_content and spend the video budget on motion. You can also skip the still: one-shot Omni I2V (client.interactions.create, video_config.task in image_to_video) is a valid call. A 1K still is ~$0.067; an 8-second Lite clip is $0.40. The seconds dominate, so you stack a still for control, not to save money.

What AI specifically gets wrong

• Treating all video models as interchangeable. They are not. Higgsfield is a camera-control layer; Omni is Google's default video call; Veo 3.1 is the generate_videos family; Kling 3.0 is price-performance; Sora 2 is a 16-20s physics model that goes away 2026-09-24. Picking by "which sounds coolest" wastes thousands of dollars.
• Calling dead IDs. veo-3.0-* and veo-2.0-* shut down 2026-06-30. imagen-4.0-* on the Gemini API died 2026-08-17. gemini-omni-flash-preview dies 2026-09-30. Sora 2 as the production default dies 2026-09-24. Omni goes through client.interactions.create, not generate_videos. First frames go through generate_content.
• Text-to-video as the default. You get one shot at composing the frame, and the model picks for you. Image-to-video gives you that frame in advance, with full control. One-shot Omni I2V is the other valid default — you don't have to generate a still first.
• Prompting "cinematic." The word is meaningless to the model. Shot vocabulary (medium close-up, push-in, rack focus) is meaningful. "Cinematic" is what someone writes when they don't know the words.
• Static cameras. A 5-second locked-off shot from any 2026 model is the new "AI smell." Audiences clock it instantly. The fix is camera motion — Higgsfield's whole pitch, or a named move in the prompt.
• Forgetting the acceptance rate. "Generate 10, pick 1" is real. Plan budget against the retake-multiplied cost, not the headline price. Lite at $3/min is $7.50/min the moment retakes hit 2.5x.

What you'll be able to do at the end

By the end of this chapter you'll have:

• A mental map of the 2026 video model lineup — which one to use for which job, by lane (default Google call, audio, camera, price, documented non-Google APIs).
• The shot vocabulary that separates AI-generated craft from AI-generated slop. Specific moves, specific lenses, specific timing.
• A routing function that picks a live model ID from a brief.
• A function that rejects shut-down IDs and date-gates Sora (2026-09-24) and gemini-omni-flash-preview (2026-09-30).
• A cost-per-finished-minute model that accounts for retake rate. Run it on your idea before you spend the money.

After this chapter, when someone says "we should make an AI video," you'll be the person in the room who can name the right model, write the shot list, and price it out before anyone opens a browser tab.