MiniMax H3 API: the cheapest Sume job is 5 s at 480p, $0.3125

Short answer

If you want the MiniMax H3 API without a separate vendor account, Sume lists it as minimax-h3. A job is 5 to 15 seconds at native 480p or 768p, and Sume bills 2K or 4K upscales only if your request asks for them (Video generation docs). The cheapest job is 5 seconds at 480p: $0.3125.

On MiniMax's own news page (read 2026-10-05), H3 is marked as new in the video models list, but that page carries no API price or availability statement in the content I could read. So the rates below come from Sume's pricing table, not from a MiniMax page.

Price by length and tier

Sume bills list times 1.25. The table uses the per-second rates in Sume's pricing table (list $0.05, $0.06, $0.13 and $0.16 for 480p, 768p, 2K and 4K).

Resolution is native, not 720p

The docs call out that 768p is first-class for this model, not 720p, and 2K and 4K are upscales of the native 768p picture. If you send resolution: "720p" expecting a standard tier, read the model's supported_resolutions first from GET /v1/videos/models.

Submit one

The request shape is the same as every Video Router model.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: h3-probe-001" \
  -d '{"model":"minimax-h3","prompt":"A slow dolly past a rainy cafe window","resolution":"480p","duration":5,"mode":"async"}'

What can go wrong

• duration: 4 is rejected on Sume even if a vendor page mentions shorter clips; the floor is 5 seconds.
• A balance below the reserve returns 402 insufficient_credits before anything runs.
• Poll GET /v1/jobs/{id}/status with backoff; do not resubmit because a poll timed out.

Which MiniMax H3 fits which job

Sume lists two H3 models. minimax-h3 is the base entry: native 480p and 768p, 5 to 15 seconds, upscales on request. minimax-h3-max is the faster 768p variant with start and end frame and reference-to-video, and a 1080p row that is a latent refinement of native 768p. For a first API test, minimax-h3 at 480p is the cheapest way to confirm the integration works end to end.

If you plan to use reference images, read the model's capabilities first. The docs say each catalog model can have a different envelope, so the field list for minimax-h3 may not match the one for minimax-h3-max.

A ten-clip test at 5 seconds and 480p costs 10 x $0.3125 = $3.125. The same ten clips at 768p cost 10 x $0.375 = $3.75. A 10-second clip at 768p is $0.75, and a 15-second one is $1.125.

Because 2K and 4K are upscales billed only if the request includes them, a request that leaves them out never pays that line. If a price looks too high, check whether the request asked for an upscale you did not need.

Before scaling, run the same prompt at 480p and 768p and compare them side by side. The 480p job costs $0.3125 and the 768p job $0.375 for five seconds, so the quality check costs under a dollar. If the 480p result is good enough for a feed thumbnail or an internal review, keep it there; the saving across a few hundred clips is real money. If you need 2K or 4K, add the upscale to the request on purpose and read the reserve before you accept it, since that is the only way those rows enter the bill.

Also record the model id with every result you store. When a vendor revises a model, the catalog entry and its rates can change, and a record of which model and resolution produced a clip lets you reproduce the cost of a past job from the ledger instead of guessing.