Skip to main content
Magic Hour exposes multiple models per endpoint. Pass the one you want in the model field, or omit it to get default, which tracks our current recommendation and changes over time. Credits are charged per generation. Video cost scales with duration, image cost with resolution and image count. Failed generations are never charged.

Video models

Available on text to video and image to video.

Choose your first model

Start with a model available on the free plan. Free-plan access still uses credits. Available on free and paid plans
  • LTX 2.5 (ltx-2.5) — our recommended starting point, especially for fast generations.
  • MiniMax H3 (minimax-h3) — a high-quality alternative.
For either model on the free plan, set resolution to 480p. Requires a paid plan
  • Kling 3.0 (kling-3.0) — our recommendation for strong quality and value.
  • Seedance 2.5 (seedance-2.5) — choose this for state-of-the-art quality.
Pass the model ID in model to select it explicitly. See the endpoint reference for supported durations and resolutions, and the credit-cost table below to compare costs.

Video model reference

default resolves to kling-3.0 on paid tiers and ltx-2.5 on the free tier. Free typically requests 480p; paid plans (Creator, Pro, Business) can request the higher resolutions a model supports, at that resolution’s credit cost. wan-3.0 requires a paid plan at every resolution. gemini-omni-1.2 requires a paid plan. Image-to-Video clips that use an end frame can be up to 10 seconds. kling-4.0 requires a paid plan and supports up to 10 API references total, including at most one video. The deprecated kling-2.5-audio and veo3.1-audio values are equivalent to passing the base model with audio: true.

Video credit costs

Credits per second of output. For text-to-video and image-to-video, multiply by end_seconds to estimate a job. For endpoints that trim an input, the clip duration is end_seconds - start_seconds. See a five-second image-to-video cost example for how credits translate under different funding arrangements. For Kling 3.0, compare the Creator, Pro and Business annual subscription example, including the full upfront payment and effective cost at each resolution and audio setting. The video formula is duration × output fps × resolution multiplier. Because output fps varies by model, two models at the same resolution can cost different amounts per second.

Video-reference pricing

Supplying a video in references can change the rate. These are total credits per output second, not additions to the base price. A listed video-reference rate takes precedence over the audio rate. For other supported combinations, use the model’s base or audio rate. A dash means no separate video-reference rate is listed. Check the endpoint reference for supported input types and model/resolution combinations.

Legacy model values

The endpoints still accept ltx-2, ltx-2.3, kling-2.5, kling-1.6, seedance for backwards compatibility, but these are no longer recommended and are absent from the tables above. New integrations should use a documented model. Existing Text-to-Video and Image-to-Video requests using gemini-omni-1.1 run on gemini-omni-1.2 at the 1.2 rates. AI video editor has its own roster: gemini-omni-1.2, wan-3.0, ltx-2.5, and kling-4.0. The default is ltx-2.5 on the free tier and gemini-omni-1.2 on paid tiers. The other models require a paid plan. On this endpoint, ltx-2.3 maps to ltx-2.5, and gemini-omni and gemini-omni-1.1 map to gemini-omni-1.2.

Image models

Available on AI image generator and AI image editor. “From” is the cost of a single image at the model’s lowest resolution. Cost depends on the model, resolution, and image count; do not assume every batch costs the single-image rate multiplied by the count. For example, nano-banana-2.1 costs 50 credits for one image at 640px or 1k, and 150 credits total for a grid of 4, 9, or 16 images. Free-plan image models are available at 640px. Larger resolutions require a paid plan even if the model is available on the free plan. Existing requests using nano-banana-2 run on nano-banana-2.1 at the 2.1 rates. For Seedream 4.5 generation, see the generator model details for available resolutions and credits.

Reference costs

These endpoints use tool-specific pricing. Some rates vary by resolution or output settings.

AI Image Upscaler

Cost depends on style.mode and scale_factor.

Duration-based endpoints

These charge by output length rather than per generation.
Video jobs return an estimate in credits_charged on the create response. The final amount is corrected once the render finishes and the true frame rate is known. Read credits_charged from the get-details response for the actual charge.

Character Replace

Credits per output second depend on model and resolution. A dash means that resolution is unsupported for that model in Character Replace. See the Character Replace guide for mode-specific duration limits and subject selection.

Estimating before you build