"AI video generator" means four different things to a developer. Here is how video-model APIs, programmatic render APIs, self-hosted open models and agent-driven studios differ, with prices and licenses I checked today.
Searching for an open source AI video generator, or any AI video generator for developers, turns up four different kinds of tool, and most roundups mix them into one list. A web search while I wrote this turned up API-vendor pages and listicles, but no guide that separates what you call, what you render from a template, what you self-host and what an agent drives. I build ReelMint, which sits in the last group, so I'll flag that where it matters and keep the comparison to things I could check today (3 November 2026). Prices below are list prices I read on each vendor's page that day; they move, so treat them as a snapshot.
The useful split is by what you hand over and what you get back.
Groups A and C make clips. Groups B and D make structured videos where text, code and charts must be legible. Picking the wrong group is the most common mistake, so decide that before you compare vendors.
These are what most people mean by "text to video for developers" or "video generation api". You call an endpoint, poll for a job, download a clip. Three things I checked that change the picture from a year ago.
First, OpenAI's Sora API is gone. OpenAI's deprecations page lists the Videos API and the `sora-2` and `sora-2-pro` models as removed on 24 September 2026, with no replacement listed. Any tutorial that starts with a Sora API call is out of date.
Second, Google's Veo previews in the Gemini API shut down on 22 October 2026. The Gemini API video docs list three Veo 3.1 preview models as shut down and recommend migrating to Gemini Omni Flash (`gemini-omni-1.1-flash`), pointing to the Gemini Enterprise Agent Platform for the GA Veo 3.1 models. The pricing page lists Omni Flash video output at $17.50 per 1M tokens and estimates roughly $0.10 per second of 720p video. I could not find a Veo per-second price there, so I am not quoting one.
Third, prices are now easy to compare on aggregators. Runway's API pricing sets a credit at $0.01 and lists `gen4.5` at 12 credits per second ($0.12/s) and `gen4_turbo` at 5 credits per second ($0.05/s); it also resells other models (for example `veo3.1` at 40 credits per second with audio, 20 without). On fal.ai, Kling v2.6 text-to-video is listed at $0.07/s and Kling v3 at $0.14/s, billed per output from prepaid credits. Replicate bills most public models by hardware time (an H100 is $0.001525 per second) and lists some video models per output second, such as `wavespeedai/wan-2.1-i2v-720p` at $0.25 per second. Luma has a public API with per-clip pricing at lumalabs.ai/api/pricing, but that page contradicted itself between its table and its FAQ when I read it, so check it yourself. Kling has a developer platform at kling.ai/dev but showed no prices to me.
What these are good for: B-roll, product shots, stylised scenes. What they are bad at: clips are short, text inside the frame is unreliable, and you pay again for every retake. For what a finished Short costs end to end, see what AI video generation actually costs rather than multiplying a per-second price by a guess.
This is the group developers underrate. If your video is data plus a layout (a changelog, a leaderboard, a code snippet, a per-user recap), you don't need a generative model at all. You describe the video and a renderer produces it, identically, every time. Searches for "programmatic video generation" land here.
The trade: you design the layout yourself, and nothing in this group writes the script or the narration for you.
If you want group A's output without per-second billing, you run the model. I read each GitHub repo and its license through the GitHub API today:
The honest cost of this group is a GPU and the time to operate it, which a license table doesn't show. "Open source" also varies a lot here, as the list above shows: two of five carry custom licenses that restrict some uses. If you are shipping a product, the license matters more than the benchmark.
Group D flips the question. Instead of you calling an API from your code, your coding agent calls the tools. ReelMint is an MCP server at `https://studio.reelmint.io/api/mcp` that Claude Desktop, Claude Code and Cursor can drive, with 130+ MCP tools covering script, scenes, voice, render and publishing. Under it is a Remotion-based render pipeline, so scenes are real components (code cards, terminal cards, chart animations, character scenes), and every generated scene stays editable in a visual editor rather than coming back as a flat file. It is bring-your-own-key: you plug in your own model and voice keys, covered on the BYOK page, and there is a free tier on the pricing page. It publishes to YouTube Shorts; Instagram Reels publishing is pending Meta App Review.
For a measured number: on the MCP path, a 60 to 90 second reel took the founder (me) 45 to 60 minutes elapsed, of which 10 to 20 minutes was hands-on instruction. That is one person, who built the product, on one path, so a newcomer will be slower. Our guide to judging MCP servers for video lists the criteria I'd apply to any of them, ours included.
What this group is: a pipeline for narrated explainers and Shorts. What it is not: a source of photoreal footage.
A decision list, in the order the questions usually resolve:
Most real projects combine two groups. A common pairing is a render API or Remotion for the structured parts and a video-model API for a few seconds of footage.
ReelMint is the wrong pick if you need photoreal people, product footage or cinematic shots: use a group A API (Runway, Luma, Kling) or a group C model. It is the wrong pick if you want a renderer you call from your own backend with a template per customer: Shotstack, json2video, Creatomate or Remotion Lambda are built for that. It is the wrong pick if you need on-premise generation with no third-party calls: that is group C. And it is aimed at vertical Shorts and explainers, so it is not the tool for long landscape video.
Not a free generative one that I could verify. The free options I found are in the render group: Shotstack's `stage` sandbox is free with a watermark, and json2video's free plan gives 600 credits with a watermark. For free generative video you can self-host an open model, where the cost moves to your GPU.
It depends on your license and hardware, and I can't honestly name one winner. On license alone, Wan 2.2 and Open-Sora are Apache-2.0 in their repos, while LTX-2 and HunyuanVideo use custom community licenses with restrictions. On recency, Wan 2.2, LTX-2 and HunyuanVideo each had commits within the last two months. Check the model card for the weights license, which can differ from the code.
Yes, with groups A, B and C. Render APIs such as Shotstack and json2video take a JSON description over HTTP, so a CI job can build a release-notes video on every tag, and Remotion Lambda does the same inside your own AWS account. Video-model APIs work the same way but add job polling and per-clip cost. Our changelog video guide shows the pattern for release notes.
Get one practical MCP-video tip when we publish the next post — no more than that.
Point your agent at a video pipeline that speaks MCP and make one small thing. Free founding-creator pilot — your keys, no watermarks.
Create your studio — free pilot