CogVideoX-5B

thudm/cogvideox-5b · by Zhipu AI (THUDM) · Diffusion transformer (DiT)

Pricing — 1 offering(s)

fal.ai

About this offering: CogVideoX-5B is an open-weight text-to-video model from Zhipu AI (THUDM). Available on fal.ai as a hosted inference endpoint at $0.20 per 10-second clip. Also available on Replicate (cuuupid/cogvideox-5b) at variable compute cost (~$0.32/run on A100 80GB). Supports video-to-video via fal.ai. MIT-licensed model weights; commercial use permitted. Effective rate ~$0.02/s.

Text-to-Video

$0.20 / clip Current 2026-06-15 → present

$0.20 per 10-second video as listed on the fal.ai CogVideoX-5B model page. Effective rate: ~$0.02/second…

$0.20 per 10-second video as listed on the fal.ai CogVideoX-5B model page. Effective rate: ~$0.02/second. Price verified from fal.ai model page (June 2026). CogVideoX-5B outputs 10 seconds at 8fps by default. Confidence marked estimated since the fal.ai pricing page was not directly accessed; price derived from fal.ai/models/fal-ai/cogvideox-5b page reference.

Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.

See how CogVideoX-5B fits into a cost-aware routing setup

Capability profile

How CogVideoX-5B rates across core capability dimensions, with the task-level evidence behind each rating.

Prompt adherence Moderate

Solid text-to-video prompt adherence for an open-weight model.

temporal consistency Moderate

Full 3D attention gives reasonable frame consistency within 10-second clips.

motion quality Moderate

Smooth motion generation; slower inference than distilled alternatives.

native audio Weak

No native audio generation.

character consistency Moderate

camera control Weak

Camera control via prompt only; no explicit camera parameter API.

clip duration ceiling Moderate

10-second clips (49 frames at 8fps).

Resolution ceiling Weak

720p output; lower resolution than closed-source alternatives.

Inference speed Weak

Slower than distilled/consistency models (LTX Video, Wan). Full 3D attention over 49 frames is compute-intensive. fal.ai cold-start latency adds additional delay. Not suitable for interactive or time-sensitive workflows.

cost efficiency Strong

video editing Moderate

V2V generation mode (via fal.ai API) accepts an existing video clip and a text prompt, outputting a new video with the input's motion and composition preserved under the prompted style — a form of style transfer on existing footage. No inpainting, mask-guided editing, or character reference exposed.

Operator guidance

Choose CogVideoX-5B when open weights and fine-tuning rights are needed, or for budget-sensitive workflows (~$0.02/s effective rate via fal.ai). Quality is competitive with other open-source models (LTX Video, Wan) but below closed-source frontier models. Inference is slower than distilled alternatives.

Use cases

Open-weights video generation with commercial-use license (MIT)
Fine-tuning and LoRA customisation workflows (weights publicly available)
Budget video inference via fal.ai or Replicate

Provenance notice

Open weights allow full inspection and audit of model behaviour
Chinese-origin research model; content policy differs from Western providers

Limitations

8fps output lower than closed-source models (typically 24fps)
720p maximum resolution
Slower inference than distilled/consistency models
No native audio generation