Google Veo 3.1 Explained (2026)

What is Google Veo 3.1?

Veo 3.1 is Google DeepMind’s flagship text-to-video and image-to-video model, released October 15, 2025. It turns a short text prompt (or a still image) into an 8-second clip with synchronized audio — dialogue, music, ambient sound — in a single generation. Earlier video AI had to be paired with separate text-to-speech or sound-effect models, so most AI video clips were silent or had obviously-grafted-on audio. Veo 3 was the model that ended that pattern, and 3.1 is the version most people will actually use.

Veo 3.1 is a closed-weights foundation model, which means you cannot download it and run it yourself. It runs only inside Google’s products and APIs: Vertex AI for enterprise, Flow for individual creators on Google AI Pro and Ultra, Google Ads for advertisers, and Gemini for casual prompting. The model card itself sits at deepmind.google/models/veo.

Three variants exist as of May 2026:

• Veo 3.1 — full model. Highest quality, highest cost.
• Veo 3.1 Fast — faster, cheaper, 70-80% of the quality on most prompts.
• Veo 3.1 Lite — cheapest, launched March 31, 2026, priced at “less than 50% the cost of Fast.”

What makes Veo 3.1 different from earlier AI video?

Four capabilities set Veo 3.1 apart from the previous generation of AI video.

Native synchronized audio. When you generate a clip of a barista pulling an espresso shot, you also hear the grinder, the steam wand, and the murmur of customers. When the clip includes a person speaking, the lip movement and the words line up. Demis Hassabis, head of Google DeepMind, framed this as “ending the silent film era” of AI video — and that framing holds up. Until Veo 3, every major AI video model produced silent clips.

Scene extension. Veo 3.1 can take an existing clip and extend it forward in time, preserving subject continuity (same person, same outfit, same camera angle) instead of restarting from zero. This is how teams produce a 30-second ad from chained generations rather than waiting for a model that natively outputs 30 seconds.

Narrative control editing. The October 15, 2025 launch added what Google calls narrative control — the ability to direct what happens at specific moments inside the clip, not just at the start. You can tell Veo “at 4 seconds, the person turns toward camera and smiles” and the model will execute that beat.

4K upscaling. Added January 2026. The base model still generates at 720p or 1080p, but Google’s upscaler can take that output to 4K — which matters for broadcast, large-format display, and post-production grading.

How much does Veo 3.1 cost?

Veo 3.1 pricing splits into API pricing (pay per second of video generated) and subscription pricing (a monthly fee with a clip allowance through Google AI Pro or Ultra).

API pricing as of May 2026:

Variant	Price per second	8-second clip
Veo 3.1 standard, with audio	$0.40 to $0.75	$3.20 to $6.00
Veo 3.1 Fast	$0.10 to $0.15	$0.80 to $1.20
Veo 3.1 Lite (720p)	$0.05	$0.40
Veo 3.1 Lite (1080p)	$0.08	$0.64

The Fast price was cut on April 7, 2026, which is why it now lands below most competitors. Google’s own announcement positioned Lite as the cheapest way to test ideas before committing to the full model.

Subscription pricing through Google’s Flow product:

• Google AI Pro — $19.99/month. Includes 90 Fast clips per month.
• Google AI Ultra — $249.99/month. Includes roughly 2,500 Fast clips per month.

For most creators, the math is: do you need to make more than 90 clips a month? If yes, you will churn through the Pro allowance and end up on Ultra or paying API rates. If no, Pro is the better starting point.

How do you actually use Veo 3.1?

Three paths, depending on what you want to do.

Casual prompting in Gemini. Open Gemini, type a prompt like “a close-up of a barista pulling an espresso shot, warm morning light, slow camera dolly forward, soft jazz in the background,” and Veo will generate the clip. This is the simplest entry point.

Structured generation in Flow. Flow is Google’s creator-focused tool that wraps Veo with a timeline, keyframe controls, and the ability to chain clips with scene extension. If you are making more than the occasional one-off, Flow is where you will spend your time.

Programmatic access via Vertex AI. For developers and teams building Veo into a product or workflow, the API lives in Google Cloud’s Vertex AI. This is how WPP runs concept testing for clients and how MNTN generates automated ad creative variations.

Prompting Veo well takes practice, but three principles travel everywhere:

1. Describe the camera, not just the scene. “Low-angle handheld” produces a very different clip from “drone overhead.” Be explicit.
2. Specify the audio. If you do not mention sound, Veo will pick something. If you do mention sound — “no music, only the crackle of fire” — you get cleaner output.
3. Use scene extension instead of chasing longer single clips. Two 8-second clips chained with scene extension look better than one model that natively outputs 16 seconds with quality degradation.

How does Veo 3.1 compare to Runway, Kling, and Seedance?

Three direct competitors as of May 2026: Runway Gen-4.5, Kling 3.0, and Seedance 2.0.

Model	Max clip	Native audio	Mid-tier API cost/sec
Veo 3.1	8 sec (16-30 sec extension Q2 2026)	Yes	$0.40-$0.75
Runway Gen-4.5	10 sec	Yes	~$0.25
Kling 3.0	15 sec	Yes (multilingual)	$0.084-$0.168
Seedance 2.0	15 sec	No	$0.14

The honest framing: Veo wins on audio, Runway wins on creator tooling, Seedance wins on price per second, and Kling wins on style versatility. There is no single best — there is a best for your specific brief.

For a deeper side-by-side, see our Sora vs Runway vs Kling comparison (being updated to add Veo and Seedance) and the broader AI Video Generation Guide for 2026.

What are Veo 3.1’s limitations?

Veo 3.1 is impressive, but four limitations matter if you are planning real work around it.

Eight seconds per generation is short. A typical YouTube ad is 15-30 seconds; a typical Instagram Reel is 60. To produce anything beyond the 8-second window today, you chain generations with scene extension and accept the small continuity drift that introduces.

Closed weights. Veo only runs inside Google’s products. You cannot fine-tune it on your brand, run it on your own GPU, or deploy it inside a private network. If those constraints matter, an open-weights model (none of which currently match Veo’s audio capability) is your only option.

Realistic people are still a gamble. Veo handles landscape, product, animal, and B-roll footage well. Realistic close-ups of human faces and dialogue-heavy scenes still get the occasional uncanny moment — wrong eye direction, lips slightly out of sync, hands with the wrong number of fingers visible at the edge of a frame.

Cost adds up fast at the standard tier. A single 8-second clip at $0.75 per second is $6. Iterating to find the right shot, which is the normal creative process, burns budget quickly. Most teams generate first at Lite or Fast for ideation, then re-roll the chosen shot at Standard.

Frequently Asked Questions

Is Veo 3.1 free?
No, but Google AI Pro at $19.99/month includes 90 Fast clips, which is the cheapest meaningful access point. For one-off experimentation, the Lite API tier is $0.05 per second — about $0.40 for an 8-second clip.

Can I use Veo 3.1 commercially?
Yes, on paid plans. Google explicitly permits commercial use of generated content from Vertex AI, Flow on paid tiers, and Google Ads. Free-tier Gemini access has more restrictive terms — read those before commercial use.

Does Veo 3.1 add a watermark?
Generated clips include SynthID, Google’s invisible watermark identifying the content as AI-generated. There is no visible watermark on paid-tier output. Free-tier Gemini output sometimes includes a visible label, which is normal for AI-content disclosure on consumer products.

How long does a generation take?
Veo 3.1 Standard runs about 30-90 seconds per 8-second clip. Veo 3.1 Fast is 10-30 seconds. Lite is comparable to Fast. Generation queue load can push these up during peak hours.

Does Veo train on my prompts?
On paid tiers (Vertex AI, Flow Pro and Ultra), Google contractually does not train on your inputs by default. On free Gemini, the default is opt-in to training, with controls under privacy settings. Read the active terms when you sign in — they update.

What about the Sora 2 shutdown — should I just use Veo now?
Yes, if your top requirement is the model staying available. OpenAI announced Sora 2’s discontinuation on March 24, 2026, with the web/mobile app closing April 26, 2026 and the API following September 24, 2026. For anyone building a workflow that depends on a model that will not be deprecated next quarter, Veo, Runway, and the Chinese models (Seedance, Kling) are the durable bets.

Sources

• Google DeepMind — Veo model page
• Google Blog — Veo 3.1 Lite announcement (March 31, 2026)
• Demis Hassabis on X — Veo 3.1 launch (October 15, 2025)
• Costgoat — Google Veo pricing breakdown
• OpenAI Help Center — Sora 2 discontinuation notice

You might also like

• The Real State of AI Video in 2026 — our Special Report Vol. 2, with the full landscape, market data, and Hollywood quotes.
• AI Video Generation Guide for Beginners (2026) — beginner walkthrough of every major AI video model.
• Every AI Model Worth Knowing in 2026 — 30+ models compared in one table.
• Sora vs Runway vs Kling — head-to-head of the leading AI video tools (being updated to add Veo and Seedance).
• AI for YouTube Creators — how creators are using Veo and other AI video models in production.
• AI Glossary — every term in plain English.