Testing without the bill mattering

### The short answer

There is no free text-to-video API. Checked, not assumed, as of September 2026:

• OpenRouter does serve video now — Seedance, Veo, Wan, Sora — but none of them are free.

There is no :free variant the way there is for its language models.

• Hugging Face gives $0.10/month of Inference Provider credit on the free tier. That is about one

five-second clip.

• Google has no free tier for Veo through the API at all. The ten free generations people mention

are in Google Vids, a consumer interface, not something you can point a pipeline at.

• Self-hosting Wan, LTX or Hunyuan genuinely is free — and needs an NVIDIA GPU with 12–24GB of

VRAM. On an ARM machine or anything without CUDA it is not an option.

### What about NVIDIA's free tier?

Asked often, because [build.nvidia.com](https://build.nvidia.com) really does hand out API credits

with no credit card and its catalogue really does list video models. It was checked properly and it

is not a way to generate directable video:

• Cosmos 3 does generate video from text, and is the one that looks like the answer. It ships as

a downloadable container to run on your own NVIDIA GPUs — it is not on the hosted API — and it

is built for physical-AI world simulation, robots and self-driving, not cinematic footage.

• Stable Video Diffusion is on the hosted API and is free, but it takes no prompt at all. It

animates a still with roughly four seconds of undirectable drift. You cannot tell it what happens

in the shot, which is the entire point of a shot.

Chaining a free text-to-image model into it was evaluated and rejected for that reason: prompting the

first frame is not directing the shot.

So the question worth answering is not "what is free" but **"what makes testing cost so little that it

stops being a consideration"**. That has a good answer.

### Most of the pipeline is already free

The cost is not spread evenly. A storyboard is one image; a rendered shot is dozens of frames of

video. They differ by roughly two orders of magnitude, and **the storyboard gate exists to exploit

exactly that**.

Pollinations FLUX needs no key at all. With nothing configured you can:

• write the script,
• build every Bible,
• generate every storyboard frame,
• assemble the timeline,
• write and translate the captions,
• export the EDL, the plan and the lookbook PDF,

…for nothing. Only the video renders cost money, and the gate means you approve a free still before

each one.

### The cheap tier, for when you do need motion

Set OPENROUTER_API_KEY and four models appear:

| Model | Per second | 5-second shot | Good for |

| --- | --- | --- | --- |

| Seedance 2.0 Mini | ~1.3 credits | ≈ $0.07 | Everything, while iterating |

| Seedance 2.0 Fast | ~4.0 credits | ≈ $0.20 | A draft that is nearly right |

| Wan 3.0 | ~4.3 credits | ≈ $0.21 | Open-weight lineage, hosted |

| Veo 3.1 Lite | ~5.0 credits | ≈ $0.25 | The cheapest native lip-sync |

A forty-shot short on Seedance Mini is under three dollars. That is the number that changes the

decision.

Select them by setting Economy quality mode — you are telling the Router a cost ceiling, not

picking a model, and it selects from whatever is live beneath it.

### Set a spend ceiling before you walk away

`

# .env.local

OPENROUTER_API_KEY=sk-or-v1-...

OPENROUTER_MAX_SPEND_USD=5

`

The ceiling is a refusal, not a warning. Past it, nothing is submitted and nothing is charged.

Two details that matter:

• It is checked against real recorded spend read back from provenance, not an in-memory counter.

A counter resets every time the dev server reloads, which during a testing session is constantly —

and a cap that forgets is not a cap.

• A warning would be a line in a log nobody reads until the invoice arrives. A testing session is

exactly when a runaway loop happens: you are trying things, something retries, and a bug you have

not found yet queues four hundred renders.

Current spend is shown at the top of Studio → Models.

### A testing order that costs almost nothing

1. 1Nothing configured. Click through every surface. The mock adapter produces real playable files,

so the pipeline is genuinely exercised — the imagery is procedural, and it is labelled as such.

1. 1Add nothing, generate stills. Pollinations is keyless. Board a whole film with real imagery

for free.

1. 1Add `OPENROUTER_API_KEY` with a $5 ceiling. Render the three or four shots you actually want

to see move. Under a dollar.

1. 1Only then consider a flagship for a hero shot, once you already know the pipeline works.

### Do not judge quality from step one

The mock renderer produces procedural SVG. It is real output through the real pipeline, and it tells

you nothing about how a model will look. It is there so you can evaluate the product without a

credential — evaluating the models is a separate exercise, and the

[Graphics Model Test](/help/graphics-model-test) is the tool for it.