Veo 3.1: the complete guide

Veo 3.1 is a diffusion-transformer video model positioned at the premium end of the market. Google's own docs describe it as a cinematic engine for high-end storytelling, and for once the marketing matches the output: golden-hour lighting, low-key scenes and high-contrast grades come out with a tonal range the other 2026 models approach but do not match. Compared to Veo 3 it improves temporal consistency, reduces flicker, and produces noticeably more natural pans, dollies and tracking shots.

• ▸Three tiers from one family: Standard (maximum fidelity), Fast (lower latency and ~4× cheaper), and Lite (draft quality for iteration).
• ▸Audio is generated with the video by default — ambience, sound effects and speech. In our UGC promo test the voice lip-synced correctly on the first attempt.
• ▸Inputs: text prompt (up to 1,024 tokens), an optional start-frame image (up to 20 MB) for image-to-video, up to 3 reference images, and negative prompts.
• ▸Generation is fully asynchronous: predictLongRunning returns an operation you poll (~every 10 s) until the video URI is ready. Typical wait: 1–6 minutes.

Veo 3.1 bills per second of output video. Cost is fully deterministic from three request parameters: tier, resolution, and duration. The official per-second rates:

• ▸Disabling audio (generate_audio: false) cuts roughly a third off any tier — about $0.13/s on Standard and $0.05/s on Fast. Worth it only when you are replacing the track in post anyway.
• ▸Billing gotcha: the docs show "$0.05 per video" next to Lite, which developers have repeatedly misread. Billing is per second — an 8-second Lite clip is $0.40, not $0.05.
• ▸Cost is charged per generated video; a request with 4 output candidates bills 4×.

From our lab: the 8-second 1080p 9:16 UGC promo on the Fast tier cost $0.96. The identical brief on Standard would have been $3.20, and honestly the Fast output was good enough to ship — our default recommendation is Fast for social work, Standard only for hero shots.

Access is through the Gemini API (generativelanguage.googleapis.com, API-key auth) or Vertex AI (IAM). The request schema:

operation = client.models.generate_videos(
    model="veo-3.1-fast-generate-preview",
    prompt=(
        "Handheld selfie video, young woman in a bright kitchen holding a serum bottle. "
        '"Okay so I have to tell you about this," she says, excited. '
        "SFX: soft room tone, morning birds outside."
    ),
    config=types.GenerateVideosConfig(
        aspect_ratio="9:16",
        resolution="1080p",
        duration_seconds=8,
        generate_audio=True,
    ),
)
# poll operation until done, then download the video URI

Requires google-genai Python SDK 1.52+. Up to 4 output videos per prompt. Output is 24 fps MP4, returned as a download URI or base64.

• ▸Veo has the strongest prompt adherence of the 2026 field — it rewards specific, structured prompts: shot type, lens feel, subject, action, lighting, then audio cues.
• ▸Write audio into the prompt explicitly with an "SFX:" line (rain, room tone, footsteps). Veo treats it as a first-class instruction, not decoration.
• ▸Dialogue in quotes is lip-synced. Keep it to one or two short lines per 8-second clip; give the speaker an emotion ("she says, excited").
• ▸Timestamp prompting works on 3.1: you can specify what happens at 0s / 3s / 6s to choreograph a beat change inside a single clip.
• ▸Use negativePrompt for recurring artifacts (extra fingers, warped text, brand logos) instead of stuffing "no X" into the main prompt.
• ▸Only 16:9 and 9:16 exist. If the deliverable is square or 4:5, generate 9:16 with safe margins and crop in post.

• ▸Wins: cinematic polish and color science, prompt adherence, speech quality, and — at the Fast tier — a surprisingly good price for 1080p with audio.
• ▸Loses on duration: the 8-second ceiling forces chaining for anything longer, and Veo chains less cleanly than Seedance's reference stack.
• ▸Loses on reference control: 1–3 images can't lock brand colors, wardrobe and framing the way Seedance's 12-input stack can — multi-scene campaigns drift.
• ▸Aspect-ratio support (16:9/9:16 only) is the most restrictive of the field.

Our verdict: Veo 3.1 Fast is the sleeper deal of 2026 — near-Standard quality at a quarter of the price, with audio included. Use Standard when a client is paying for the final 10% of polish on a hero shot, Lite to iterate, and switch to Seedance 2.0 the moment the job needs continuity across scenes or clips longer than 8 seconds.