Which AI video model should you use? A 2026 developer's guide

There's a large catalog of video and image models on SamAutomation, all reachable through one API key. That's a lot of choice — and the honest answer to "which model should I use?" is "it depends on the job." This guide maps the models to the jobs, so you can pick fast.

Two things up front. First, the core product is still JSON-to-video — deterministic rendering from a JSON payload, built for automation. AI generation is what you reach for when you don't have source footage. Second, every model below runs through the same lifecycle: POST /api/ai/jobs with a model_id and options, then poll GET /api/ai/jobs/{id} until it's done. Swapping models is a one-field change. See the API docs for the full contract.

Text to video — start from a prompt

This is the core generation tier. You describe a scene, you get a clip.

• Veo 3.1 (Google DeepMind) — cinematic quality with synchronized native audio in one pass. There's a cost-efficient Lite tier, a low-latency Fast tier, and a top-quality tier. Reach for it when you want dialogue, ambient sound and music generated together with the picture.
• Kling 3.0 (Kuaishou) — director-grade generation with multi-shot storyboarding and native audio, plus a faster, lower-cost Turbo tier for quick iteration.
• Wan 2.7 (Alibaba) — a full suite (text, image, reference and instruction-edit) with native audio; good when one model needs to cover generation and editing.

Rule of thumb: need audio baked in and cinematic adherence? Veo 3.1. Need multi-shot or a cheap fast tier? Kling.

Multimodal & reference — steer with more than text

When a prompt alone isn't enough — you have a character, a product, a reference clip or a voice to match.

• Seedance 2.0 (ByteDance) — combine text, image, video and audio references in a single call, with native in-pass audio and multi-shot cuts. A Fast tier trades a little quality for speed.
• Gemini Omni (Google) — any-input-to-video with voice and character-consistency workflows; useful for personalized presenters where a character has to stay recognizable.
• Kling 3.0 Motion Control — motion transfer: feed a reference video and a character image, and it re-animates your character with the reference's motion. Ideal for a brand mascot performing a specific gesture.

Rule of thumb: keeping a character consistent across shots → Seedance 2.0 or Gemini Omni. Copying a specific movement → Motion Control.

Avatar & lip-sync — make someone talk

Two distinct paradigms here — don't confuse them.

• OmniHuman 1.5 (ByteDance) — animate a still: one portrait plus a speech track becomes an expressive talking avatar with context-aware gesture and head motion.
• Video Lip Sync — re-sync existing footage: take a real talking-head video and align the mouth to new target audio. This is the dubbing/localization path — update the script or swap the language without a reshoot.

Rule of thumb: starting from a photo → OmniHuman. Starting from existing video → Lip Sync.

Video editing — change a clip you already have

• Wan 2.7 and HappyHorse 1.1 (Alibaba) — instruction-based edits, reference-to-video and extend operations on existing footage. HappyHorse adds native multilingual lip-sync, so it doubles as a talking-character tool.

See the whole catalog

Lite, Fast, Pro — pick the right tier

Several families ship in tiers, and the cheapest variant is usually the smart default while you're still iterating:

• Lite / Mini — a fraction of the credits. Use it to nail the prompt, framing and motion before you spend on quality.
• Fast — quick drafts and previews at lower latency.
• Quality / Pro — the top tier, for the final production render.

For example, Veo 3.1 runs Lite at ~40 credits, Fast at ~100 and Quality at ~220 — so you can prototype for roughly 5× less, then switch the model_id to the Quality tier for the hero render. Seedance 2.0 works the same way (Mini → Fast → full). The rule of thumb: test cheap, ship premium — same call, same payload, only the model_id changes.

How pricing works, so cost never surprises you

Everything is priced in credits, and credits scale with the work: a short 720p clip costs less than a long 1080p one on the same model. The number you see in GET /api/ai/models for any option combination is the number you pay — query it before you call. Your monthly plan includes credits across the catalog, and failed jobs are refunded automatically. Compare tiers on the pricing page.

One call to rule them all

Because the catalog is data-driven, new models appear in GET /api/ai/models as we add them — your integration doesn't change. Pick a model_id from the table, send a job, poll for the result, store the URL. If you're combining AI clips with deterministic rendering, hand the output URLs straight to JSON-to-video. Start in the AI API docs.