Test a faster-and-cheaper claim with your own timings and usage.cost

To test a claim like faster and cheaper, run the same prompt through both models, record wall-clock seconds and the billed cost for each run, and compute two ratios. Luma's news page says Ray3.14 (January 2026) is native 1080p, 4x faster and 3x cheaper. The page text read here does not say what it is compared against, so only your own runs can tell you what that means for your clips.

What to record

On Sume, a job returns usage.cost, the billable amount, and its status moves through pending, in_progress and completed (or failed or cancelled). Time from your submit call to the first completed status, and read usage.cost from the finished job. See jobs and results and the video docs.

• Same prompt, duration, resolution and aspect ratio for both models.
• At least five runs per model; queue time varies.
• Use medians, not means, so one stuck job does not decide the result.
• Count failed jobs as a separate column.

The ratio script

Feed it lists of seconds and costs from your own log. It prints how many times faster and cheaper the second model was, using medians.

from statistics import median

def ratios(a_secs, a_costs, b_secs, b_costs):
    if not (a_secs and b_secs and a_costs and b_costs):
        raise ValueError('need at least one run per model')
    faster = median(a_secs) / median(b_secs)
    cheaper = median(a_costs) / median(b_costs)
    return faster, cheaper

if __name__ == '__main__':
    # replace these with your own measured runs
    a_secs, a_costs = [100.0, 120.0, 110.0], [1.0, 1.0, 1.0]
    b_secs, b_costs = [50.0, 60.0, 55.0], [0.5, 0.5, 0.5]
    f, c = ratios(a_secs, a_costs, b_secs, b_costs)
    print(f'{f:.2f}x faster, {c:.2f}x cheaper')

Reading the result

The numbers in the demo block are placeholders to show the output shape, not measurements. A ratio above 1 means the second model won on that axis. If speed improves but cost per clip does not, say so; a claim that holds on one axis is still useful when the other axis is not your constraint.