Caption once after stitching: four 15 s clips cost $0.30, not $0.90

Burn captions once, on the stitched Reel. Four 15-second clips captioned one by one cost 4 x $0.20 = $0.80 and still need a $0.10 Timeline render to join them, $0.90 in all. Stitch first and caption the 60-second result and the bill is $0.10 + $0.20 = $0.30.

First Draft on an iPhone assembles your selected clips into one starting Reel (read 2026-10-09); the same order of work applies in the API: assemble, then caption the finished cut.

The two orders side by side

The caption price is flat per accepted job and is stated for videos of up to 60 seconds, so the saving only holds while the stitched Reel stays inside that figure. Timeline bills $0.10 per ceil output minute, so 60 s is one minute.

Both steps in Python

The helper posts a job, polls /v1/jobs/:id/status until it is terminal and returns the result. The render result carries video_url; the caption request takes it as video_url. script_text is optional; leaving it out burns the speech-to-text words.

import os, time, requests
API = "https://api.sume.com"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def run(path, body, key):
    r = requests.post(API + path, json=body, timeout=60,
                      headers={**H, "Idempotency-Key": key})
    r.raise_for_status()
    job = r.json()["data"]["request_id"]
    while True:
        st = requests.get(f"{API}/v1/jobs/{job}/status", headers=H, timeout=30)
        st = st.json()["data"]
        if st["sume_status"] in ("completed", "failed", "canceled"):
            break
        time.sleep(st.get("next_poll_after_seconds") or 3)
    if st["sume_status"] != "completed":
        raise RuntimeError(f"{path} ended {st['sume_status']}")
    res = requests.get(f"{API}/v1/jobs/{job}/result", headers=H, timeout=30)
    return res.json()["data"]["result"]

cut = run("/v1/timeline-1.0/render", timeline_body, "reel-42-render")
final = run("/v1/video-captions",
            {"video_url": cut["video_url"], "style": "punch"}, "reel-42-caps")

Caveats

• Caption words, cues and segments timings are capped at 60 s in the schema, matching the 60-second price note. For a longer Reel, plan one caption job per 60 s.
• A stitched render with no speech in its audio spine fails with caption_no_speech; send cues instead.
• punch and tiktok-green ignore design overrides; pick slam or a Hangul style when you need to tune colors.
• Keep the same Idempotency-Key on a retry so the caption job is not billed twice.