The Limits Are Measured Before The Model Lands

The limits were measured on live API probes rather than read off a model card, and the probe corrected 2 things an earlier internal note had wrong. The first is the resolution string. 768P and 1080P are values. Lowercase 768p is an error, and nothing in the answer tells you the case was the problem. The second is the endpoint. The H3 family does not live on the endpoint the rest of the family uses, which is a different class of mistake from a wrong parameter. The probability that a limit read off a model card survives the first live call sits near 1 in 4, against near 24 in 25 for the limit that was probed. A wrong resolution string fails on the first submit, and that is cheap. A wrong endpoint fails after the integration is written, and that is not. The file is 333 words of measured limits. Every number in it came back from a call rather than from a page.

3 calls carry the entire workflow: submit, poll, retrieve. Submit is 1 POST to the video generation path, and the answer is a task id. Poll is 1 GET against the query path with that task id, and the answer is a status: Preparing, Processing, Success or Fail. Retrieve is 1 GET against the files path with the file id, and the answer is a download URL. 1 bearer header covers all 3, and no client library is needed. A short script holds submit, poll and download, and that script is the whole integration. Everything else in the file is a limit on what those 3 calls will accept. The workflow is small, and the boundary around it is where the hours go.

Hailuo-2.3 does text to video and image to video at 6 seconds or 10 seconds. There is no 15 second setting. 10 seconds at 1080P does not exist either. The 2 durations and the 2 resolutions are not free combinations, and the empty square is the one a schedule usually reaches for. Image to video takes a public first frame URL, which is the next step of this file and the one that decides an upload. S2V-01 accepts 15 seconds. It accepts them for a subject with a clear character or face, passed as a subject reference against the image. A faceless back-view figure returns an unexpected error 1000 rather than a render, and the code says nothing about the subject being the reason. MiniMax-H3 sits on a different endpoint and is gated by the plan. Ask for it on the coding key and the answer is 2013. 3 models, 3 different reasons a call fails, and not 1 of the 3 reasons is a bug.

Error 2013 reads like a missing model and it is a limit on the plan. Error 1000 on a faceless subject reads like a bad request and it is a model that needs a face. A lowercase resolution string reads like a malformed parameter and it is a case-sensitive value. 3 failures, 3 wrong readings, and each wrong reading sends the work somewhere else: a support ticket, a rewritten prompt, a rebuilt integration. The reading that costs the most is the first 1. A missing model invites a switch of vendor, and the switch is unnecessary. Read every code as a statement about the account or the model rather than about the request, and 3 dead ends close in 1 pass.

The limits were measured before the model landed. That order is why the table holds.

The first frame must be a public URL. A file sitting on the machine that submits the call is not a remote input. A provider CDN URL works directly, and a plain file host works as the fallback when the still was rendered somewhere else. The failure shape is quiet. A local path returns an error that reads like a bad file rather than a bad address, and the file is perfectly good. So the still has to be reachable from outside the machine that made it, which is 1 upload step and it is not optional. The same rule holds far outside the API. In the building at Lehman Brothers for the 2003 blackout, the input that had been checked was the input that was trusted, and 200 wood-gasification machines built across the UK and Europe were specified the same way: the input that is written down is the input that survives the far end.

1 recipe closes the file: render the still first, gate it, then animate it as the first frame. The still is 16:9 and 1 image per call. The animation carries it at 10 seconds, 1364 by 768, at 24 frames a second. That is the longest shot the coding plan key allows, and it is a measured ceiling rather than a recommendation. The 15 second shot needs 1 of 3 routes: a clear character subject on S2V-01, an H3 plan upgrade, or a dialogue model with no image conditioning at all. Notice the order. The still is measured, gated and approved before the animation starts, so an image defect is caught at the price of an image rather than at the price of a video. Companies built across 12 countries, a EUR 75 million industrial group restructured, 210 energy systems across Africa and Asia, and 200 wood-gasification machines across the UK and Europe all ran on the measured version of the limit rather than the remembered 1.