Open source video model API prices

Which video models are open-weight, and which are API-only

The same brand often sells an open model and a closed one side by side. MiniMax and Alibaba publish weights for one version and keep the newer ones behind their APIs.

That split matters when you pick an API. A provider selling “Wan” may be selling closed Wan 3.0, and one selling “MiniMax H3” may be selling H3-Max, which has no weights.

Which provider sells which open model

The matrix covers eight API providers; Runpod’s Public Endpoints add Wan 2.2 in the Wan tables. Novita AI sells the H3 family only, and Alibaba sells Wan only.

Of the eight, WaveSpeedAI is the only one with a readable price for all four families.

MiniMax H3 API pricing per second, provider by provider

MiniMax-H3 costs from $0.30 for 5 seconds at 768P, on fal. MiniMax’s own API lists $0.08 a second at 768P, and Replicate and WaveSpeedAI match it at list price. Novita’s rate is $0.08 a second.

The route decides which document governs the clip. MiniMax’s API is the official route; WaveSpeedAI runs its own copy of the open weights.

Novita’s rate was read on 12 November 2026; the other rows on 14 November 2026.

WaveSpeedAI’s model page shows the open-weights edition at 50% off on its Generate button, with no end date shown. At that discount, the 5-second 768P clip costs $0.20.

Extras bill on top. MiniMax and Novita charge $0.04 for each input image beyond five, and bill input video by the second when you edit an existing clip (MiniMax pricing).

Other resolutions differ by provider.

Together lists MiniMax-H3 at 2K only; its page does not state the billing unit.

H3-Max, H3-Highspeed, H3-Lite and H3-Max-Turbo have no published weights; MiniMax’s own rates for them sit in Hailuo and MiniMax pricing. Novita’s full rate card is on Novita AI pricing, and its sign-up credit, which is Agent Sandbox credit, in the Novita AI promo code guide.

Wan 2.2 API pricing

Wan 2.2 comes in two sizes. The 5B model (5 billion parameters) runs on smaller cards; A14B is the bigger model, a pair of 14B experts. Their prices sit in separate tables because they are different models.

DeepInfra’s figure is for FastWan, FastVideo’s three-step distillation of Wan 2.2 5B.

fal’s standard A14B counts video seconds at 16 frames a second, so its default 81 frames are 5.06 seconds. Its page does not state how it rounds. Alibaba’s page does not say whether its wan2.2 flash and plus models are the published weights. Runpod’s Public Endpoints need no GPU set-up.

Closed Wan for reference: Wan 2.5 to 3.0 have no weights. Alibaba’s wan3.0-video lists $0.10 a second at 720p in Singapore and $0.083 in the US (Virginia) region.

The Singapore rate was read on 13 November 2026 and the US rate on 7 November 2026.

LTX-2 API pricing

LTX-2 is Lightricks’ open model that makes video and sound in one pass. These are the weights; Lightricks’ LTX Studio app is a separate product.

LTX-2.5 costs $0.09 a second at 720p on both fal (lightricks/ltx-2.5/text-to-video/fast) and DeepInfra (Lightricks/LTX-2.5-Fast). Both have a 6-second minimum, so the shortest clip costs $0.54.

HunyuanVideo API

HunyuanVideo 1.5 is Tencent’s open model. WaveSpeedAI sells it at $0.04 a second at 720p, so a 5-second clip costs $0.20 (wavespeed-ai/hunyuan-video-1.5/text-to-video).

fal lists two 1.5 endpoints, but their pages returned an error on 13 November 2026, so fal’s price is unverified.

Replicate sells only the 2024 version, billed by GPU time; its page puts a run at about $2.55.

Can you use these clips in ads? What each licence says

For Wan 2.2, the model licence allows it: Apache 2.0 sets no territory or revenue limit on clips. The other three licences set territory or revenue limits, and each API’s own terms still apply.

Licence

Licence The models in this repository are licensed under the Apache 2.0 License.

Outputs We claim no rights over the your generated contents, granting you the freedom to use them.

5B model The TI2V-5B model can also run on consumer-grade graphics cards like 4090.

The H3 Community License (2 August 2026) defines its Excluded Territories as “the European Union, the United Kingdom, the Republic of Korea and the United States of America” (definition 5). Clause V.4 then reads: “You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.”

Definition 5 and section V.4 (dated 2 August 2026)

5. Excluded Territories means the European Union, the United Kingdom, the Republic of Korea and the United States of America.

V.4 You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.

V.4 binds whoever runs the weights, so it reaches WaveSpeedAI’s self-hosted edition directly. Clause V.2 makes such a host bind its users to terms “at least as protective as the use restrictions in this Section V”, the territory clause included. The H3 licence question for running it yourself is on the Hailuo profile.

The official API route runs under MiniMax’s API terms, a separate document from the H3 licence. The H3 licence does not say whether its territory limit follows clips made through that API.

HunyuanVideo 1.5’s licence sets its Territory as worldwide “excluding the territory of the European Union, United Kingdom and South Korea”. Clause 5(c) bars any use of “Output or results” outside it. The US is inside its Territory, unlike H3’s.

Territory (definition l) and section 5(c)

Territory shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea.

5(c) You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan Works, Output or results of the Tencent Hunyuan Works outside the Territory.

LTX-2.5 sits under the LTX-2.x Community License, dated 11 August 2026 and amended 29 October 2026.

Its section 2.1 says entities with annual revenues of at least $10,000,000 “are required to obtain a paid license”.

Section 2.1, grant of licence (LICENSE-2_x, 11 August 2026, amended 29 October 2026)

Entities with annual revenues of at least $10,000,000 (the “Commercial Entities”) are required to obtain a paid license for any use (excluding use solely for a Non-Commercial Purpose as set forth in Section 2.2) of LTX-2.x and Derivatives of LTX-2.x.

LTX-2.3 sits under the older LTX-2 Community License of 5 January 2026, whose section 2 sets the same revenue line for a paid licence.

Section 2.1, grant of rights (dated 5 January 2026)

Entities with annual revenues of at least $10,000,000 (the “Commercial Entities”) are required to obtain a paid commercial use license in order to use LTX-2 and Derivatives of LTX-2.

Running Wan 2.2 and LTX-2 on your own card is covered in whether Wan 2.2 and LTX-2 clips can run in ads.

Providers’ own terms add a layer. Replicate grants you its rights in output “subject to any Third Party Terms (as determined by the Models you use)”. Together says you “exclusively own” output. fal’s and WaveSpeedAI’s terms contain no output ownership grant, and WaveSpeedAI bars “generating political advertisements”. Novita’s terms, section 10, neither grant nor limit commercial use of output.

For HunyuanVideo and LTX, no provider page says whether the model’s licence also reaches you as that provider’s customer.

API or a rented GPU?

Per-clip API prices win when you make a few clips, or need a model you cannot run yourself, such as closed Wan 3.0. You pay only for what you generate, with no set-up.

A rented graphics card can cost less per clip at volume on open models. The sum for one Wan 2.2 or LTX-2 clip is in what one clip costs on Runpod. An RTX 4090 costs $0.33 an hour on demand on Novita and $0.34 an hour on Runpod’s Community Cloud, billed by the second on both.

Minimax H3 vs. LTX 2.3 vs. Wan 2.2Videcool

The two clouds meet in Novita vs Runpod, and every card’s rate sits in the cloud GPU price table. The AI video hub compares app plans for the same models.

Open video API questions, answered in a line

Is MiniMax H3 open source?

No. Its weights are public, but under a community licence that excludes the US, the EU, the UK and South Korea.

Is Wan 3.0 open source?

No. Wan 3.0 has no weights on Wan-AI’s Hugging Face account and is sold by API only. Wan 2.2, under Apache 2.0, is the newest open Wan.

Which provider sells MiniMax H3 for the least at list price?

fal, at $0.30 for a 5-second 768P clip (fal model page). WaveSpeedAI’s open-weights edition costs less only at its undated 50% discount.

Do these APIs give free credits?

WaveSpeedAI gives eligible new accounts $1. Alibaba gives free quota in its Singapore region only. Novita’s sign-up credit is for Agent Sandbox. fal, Replicate, Together, DeepInfra and MiniMax state no figure on the pages we read.