Fix one wrong number: Claude Motion edit vs a Sume Format retake

Claude Motion edits a number in place. In a Sume Format run a changed word or number means a new production; a look change is a retry at about 1/20 the cost.

5 min readSume
All posts

When a number in the video is wrong, Claude Motion lets you change it in place: Anthropic's announcement says you can edit any word, number or timing in the editor or by asking Claude, because the animation is code. In a Sume Format run the rule is the opposite. A retry is a new take of a scene, so if the words, the host or the product change you start a new production; you retry only when the look changes.

The two products make different things, so the right fix differs. Motion makes an animation from code. A Sume Format generates footage, voice and assembly, where each scene is a generation.

Which fix applies

Fix routes from the Claude announcement and Sume Format docs, read 2026-10-09
What is wrongClaude MotionSume Format run
A number or word on screenEdit it in placeStart a new production; a retry regenerates the scene
Timing of one beatChange the timingRetry the scene with previous_run_id
The look of one sceneAsk Claude to redo itRetry the scene with previous_run_id
The host or the productNot applicable: no generated peopleStart a new production with new scene ids
Need a new fileDownload the MP4 againThe retry is a new run with its own id and receipt

How a Sume retry works

Send a new POST …/runs with previous_run_id naming the run that finished. The run must be continuable: its receipt shows a thread_id that is not null, and the run completed or left artifacts. The new run has a new id, its own spend cap and its own single webhook. The original run never changes.

Bind the same output_schema on every turn, because it is per run and not inherited. Name previous_run_id, not a thread id, since a thread id returns 400 unknown_parameter.

What a retry costs

The Format cookbook says measured production retries cost approximately one twentieth of the first run's spend, and tells you to always send a cap. As an illustration only, if a first run billed $12.00, a retry of about one twentieth would be 12.00 / 20 = $0.60. The example in the docs sets the retry cap at 8, so the cap is a ceiling you choose, not a prediction.

A retry regenerates all the generative parts of that scene. It is a new take, not a re-encode, so the same instruction can produce a different picture.

Read the receipt after a retry: usage.billable_amount_usd_micros is the spend of that run alone, while artifacts[] on a continued run lists all the media the full conversation generated.

Refusals to expect

If the number that is wrong came from your data, correct the data and run again from the start. Do not ask a retry to rewrite copy that the scene was built around.

  • 404 previous_run_not_found: the id is unknown or belongs to another owner.
  • 400 previous_run_format_mismatch: the run started on a different Format.
  • 409 previous_run_not_terminal: the run is still going; poll it and call again.
  • 400 previous_run_not_resumable: there is nothing to continue, so start a new run.

Sources

Related posts

More in Formats

All Formats posts

Written by Sume