Google Veo has quietly become the video model that developers ask about most. The current lineup spans Veo 2, Veo 3, and the newer Veo 3.1 variants, and the quality jump between generations is substantial; if you want the full picture of what the latest release can do, our Google Veo 3.1 overview breaks down the model family in detail. What trips people up is not the model itself. It is the access path.
Unlike a lot of AI labs, Google does not give Veo a standalone public endpoint with a simple API key signup. There are two official routes, a handful of third-party providers, and a pricing model that depends heavily on which door you walk through. Pick the wrong one and you end up wrestling with cloud IAM roles for an afternoon when all you wanted was an eight second clip.
This guide covers every practical way to call Veo from code in 2026: the Gemini API route, the Vertex AI route, and the aggregator platforms that wrap both. If you are still deciding whether Veo is the right model for your project at all, the Sora 2 vs Google Veo 3 comparison is a good place to calibrate expectations first.
The Two Official Routes: Gemini API vs Vertex AI
Google serves Veo through two distinct surfaces, and the difference matters more than the documentation suggests. The Gemini API (via Google AI Studio) is the fast path: you create an API key in minutes, no cloud project ceremony required, and you can start generating. Vertex AI is the enterprise path: full Google Cloud integration, regional control, quota management, and consumption-based billing. Both serve the same underlying models, which consistently rank near the top in our AI video generator comparisons.
Here is the practical breakdown:
- Gemini API: best for prototypes, indie projects, and anything where you want an API key today. Authentication is a single header. Veo 3.1 and Veo 3.1 Fast are both available.
- Vertex AI: best for production systems inside Google Cloud. You get regional endpoints, IAM-based auth, audit logging, and access to the full model lineup including Veo 2 and the cost-efficient Veo 3.1 Lite preview.
- Third-party providers: best when Veo is one model among several in your stack and you do not want to manage Google credentials at all.
Most solo developers should start with the Gemini API and only graduate to Vertex AI when compliance or scale demands it. Teams that already orchestrate several models through one interface tend to skip both and route everything through a unified gateway instead, a pattern we covered in how to build AI workflows with an API.
Setting Up Gemini API Access
The Gemini API route takes about five minutes end to end, roughly the same setup effort as calling FLUX 2 from code. No billing account is strictly required to get a key, though Veo generation itself is a paid feature.
- Go to Google AI Studio and sign in with any Google account.
- Click "Get API key" and create a key tied to a new or existing project.
- Enable billing on that project; Veo requests fail without it.
- Store the key as an environment variable (
GEMINI_API_KEY), never in source code. - Confirm model availability: as of mid-2026 the key models are
veo-3.1-generateandveo-3.1-fast-generate.
That is genuinely all the setup. Compare that to the credential dance some other model providers require and it is refreshing; we documented a similar lightweight flow in our FLUX Pro API pricing and code examples guide, and Google's developer experience here is comparable.
Making Your First Veo Request
Video generation is asynchronous everywhere Veo is served. You submit a job, receive an operation ID, and poll until the video is ready. A minimal request via the Gemini API looks like this:
curl -X POST \
"https://generativelanguage.googleapis.com/v1beta/models/veo-3.1-generate:predictLongRunning" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"instances": [{"prompt": "A drone shot of a coastal village at golden hour"}],
"parameters": {"aspectRatio": "16:9", "resolution": "1080p"}
}'
The response contains an operation name rather than a video. Generation typically takes one to three minutes depending on duration and load. This async pattern is identical to how other major video APIs behave; if you have followed our guide on generating videos with Kling AI via API, the polling loop will feel familiar.
curl -s \
"https://generativelanguage.googleapis.com/v1beta/$OPERATION_NAME" \
-H "x-goog-api-key: $GEMINI_API_KEY"
When done flips to true, the response includes a download URI for the generated MP4. On Vertex AI the endpoint shape differs (projects/{id}/locations/{region}/publishers/google/models/veo-3.1-generate:predictLongRunning with a Bearer token from gcloud auth print-access-token) but the request body and polling flow are the same.
What Veo Access Actually Costs
Veo pricing is usage-based on both official routes and is calculated per second of generated video, with the rate depending on model variant, resolution, and whether audio generation is enabled. Veo 3.1 at 1080p with audio sits at the premium end; Veo 3.1 Fast costs meaningfully less per second, and the Veo 3.1 Lite preview on Vertex AI is the budget option. The cost structure rewards the same discipline as image APIs, where batching and variant selection drive the real savings, something we explored in how to run batch image generation via API.
Three habits keep bills predictable. Generate at 720p while iterating and reserve 1080p for finals. Use the Fast variant for prompt exploration. And set a hard budget alert on the project before your first request, not after your first invoice.
Prompt quality is the other hidden cost lever. A weak prompt means regenerating clips three or four times, which multiplies spend directly. Spending twenty minutes with a structured prompting approach like the one in our Veo 3.1 prompt guide routinely cuts generation counts in half.
Third-Party Providers and Workflow Platforms
If Veo is just one ingredient in a larger pipeline, managing Google credentials alongside three other provider accounts gets old fast. This is where aggregators and orchestration layers earn their keep: one API key, one billing relationship, multiple models behind it. We compared the major options in our roundup of AI orchestration APIs for production apps, and the category has matured considerably this year.
For visual-first teams, the Wireflow platform takes a node-based approach: you wire Veo into a canvas alongside image, audio, and text models, then trigger the whole workflow through a single REST call. That means a product team can chain an image generation step into a Veo image-to-video step without writing the glue code for either provider's auth or polling logic.
Replicate and fal.ai also serve Veo variants with simple token auth, and both are solid choices when you need raw model access rather than multi-step orchestration. The tradeoff is that chaining models across providers still leaves the pipeline plumbing to you, which is exactly the work that REST-based pipeline patterns are designed to absorb; our walkthrough on building AI pipelines with REST APIs covers that architecture in depth.
Common Pitfalls
A few failure modes show up constantly in developer forums:
- Calling a synchronous endpoint. Veo only exposes long-running operations. If your client times out at 30 seconds, the problem is your HTTP timeout, not the API.
- Skipping billing setup. API keys without an attached billing account return permission errors that look like auth bugs.
- Wrong region on Vertex AI. Veo is served from a limited set of regions; requests to unsupported regions fail with a model-not-found error.
- Ignoring content filtering. Prompts involving real people or sensitive subjects get rejected at submission time. Rework the prompt rather than retrying.
- Treating prompts as an afterthought. Veo responds strongly to camera language and shot structure. Generic prompts produce generic footage.
The last point deserves emphasis. The gap between a mediocre Veo clip and a striking one is usually the prompt, not the parameters, and the techniques in our Veo 3 prompt collection apply unchanged to 3.1.
FAQ
Is there a free way to access the Veo API? No. Veo generation requires billing on every official route. Google AI Studio occasionally offers limited free generation in its web UI, but programmatic access is paid from the first request.
Which is cheaper, the Gemini API or Vertex AI? Per-second rates are broadly similar; the difference is overhead. Vertex AI adds cloud infrastructure controls that matter at scale but cost engineering time to set up. For benchmarking Veo's value against alternatives, our Veo 3 vs Seedance comparison includes cost-per-clip context.
Does Veo support image-to-video through the API? Yes. Veo 3.1 accepts a reference image alongside the text prompt, and it also supports first-and-last-frame interpolation. Animating a still image is one of the most common production uses.
How long can generated videos be? Standard generations run four to eight seconds. Veo 3.1 supports extending clips in segments, which is how creators assemble longer sequences; for narrative work most teams stitch clips in an editor, a workflow covered in our guide to turning text into video with AI.
Can I use Veo output commercially? Yes, subject to Google's generative AI terms. Outputs include SynthID watermarking, which is invisible but detectable, so plan disclosure accordingly for client work.
Do I need Google Cloud experience to use Veo? Not for the Gemini API route, which is a plain REST key. Vertex AI assumes working knowledge of Google Cloud projects and IAM. If neither appeals, aggregator platforms remove the Google-specific setup entirely, as do the no-code builders in our no-code AI workflow roundup.
Wrapping Up
Veo access in 2026 comes down to matching the route to the project. Prototypes and solo builds belong on the Gemini API, where a key and a curl command get you generating within the hour. Production systems with cloud governance requirements belong on Vertex AI. And multi-model products, where Veo sits inside a larger generation pipeline, increasingly belong behind an AI workflow automation platform that handles the auth, polling, and model chaining as a single managed surface.
Whichever door you choose, the model on the other side is the same. Start with the Fast variant, keep your prompts specific, and let the budget alerts do their job.