{
  "frame": {
    "model": "openai/gpt-image-2",
    "model_with_references": "openai/gpt-image-2/edit",
    "input": {
      "quality": "medium",
      "num_images": 1,
      "output_format": "jpeg"
    },
    "image_sizes": {
      "16:9": {
        "width": 1280,
        "height": 720
      },
      "9:16": {
        "width": 720,
        "height": 1280
      }
    },
    "qualities": {
      "low": "rough board: composition, order and light; the frame costs about an eighth of medium, its references the same",
      "medium": "clean board (default): clear enough to approve a film from",
      "high": "film keyframe quality; leave this to the film Skill, which re-renders each keyframe at 1536 x 864 or 864 x 1536"
    },
    "max_references": 3,
    "approx_cost": {
      "per_1024_square": {
        "low": 0.006,
        "medium": 0.05,
        "high": 0.2
      },
      "per_reference_image": 0.01,
      "note": "GPT Image 2 is token-billed (image output $30 per 1M tokens, image input $8 per 1M, text $5 in and $10 out per 1M). fal's own per-image prices (model page, read 2026-10-18): 1024 x 768 low $0.005, medium $0.037; 1024 x 1024 low $0.006, medium $0.053; 1024 x 1536 low $0.005, medium $0.042. The scripts scale the per-1024 figures by pixel count, so a 1280 x 720 frame is about $0.044 at medium and $0.005 at low, plus about $0.01 per reference image (input image tokens). fal returns no token count with the image; the fal dashboard shows the real charge."
    }
  },
  "defaults": {
    "aspect": "16:9",
    "board": "16:9",
    "shots": 6,
    "length_seconds": 30,
    "final_resolution": "720p",
    "quality": "medium",
    "board_sizes": {
      "16:9": {
        "width": 2560,
        "height": 1440
      },
      "9:16": {
        "width": 1440,
        "height": 2560
      }
    },
    "preview_max_width": 2048
  },
  "cost_note": "A 6-frame board at medium is about $0.34 on fal. A low-quality rough pass first costs about $0.11, and the clean board after it about $0.40 (each frame also carries its rough frame), so about $0.52 in all; the OSTRA sample spent $0.53 by the lab's estimates, with no re-rolls. build-requests.mjs --stage estimate prints the total for the real shot list; quote it and get one yes before the first paid call. This Skill makes no video: price-table.mjs prints what each video route would charge for the planned shots."
}