minimax-h3 vs minimax-h3-max on Sume: price per clip at 768p and 1080p

At 768p, Sume charges $0.10 per second for minimax-h3-max and $0.075 for minimax-h3, so Max costs 1.33x as much (0.10 / 0.075). Only Max lists 1080p, at $0.20 per second. Both take 5 to 15 seconds. Prices are provider list x 1.25 from the repo pricing table, as of 2026-10-08; Sume's docs call Max the faster 768p variant.

Per-second and per-clip prices

Per-clip = seconds x the per-second price.

What the docs say separates them

Both are in the same catalog and reached through the same endpoint; the differences below are from Sume's video docs.

• minimax-h3: 5 to 15 seconds at native 480p or 768p; Sume bills 2K or 4K upscales if the request includes them.
• minimax-h3-max: the faster 768p variant, with text-to-video, first and last frame, and reference-to-video at 480p, 768p and 1080p; 1080p is a latent refinement from native 768p.
• Both report image, video and audio input references; Sume's docs describe native stereo audio for minimax-h3-max.

When the premium is worth it

The extra $0.025 per second at 768p adds $0.125 to a five-second clip (5 x ($0.10 - $0.075)). On 200 clips that is $25.00 a month. If Max's speed or its 1080p output saves a re-run or a separate upscale, it can pay for itself; if you draft at 480p, both ids cost the same $0.0625 per second.

A cheap loop is to draft on minimax-h3 at 480p, $0.3125 for five seconds, and move the keepers to a higher tier.

Open weights and the hosted ids

MiniMax's H3 page says the open weights generate natively at 768p and that 2K output uses API-only stages. Sume does not claim parity with local weights, and its docs do not list a 2K tier for minimax-h3-max. The H3 license excludes the EU, UK, US and South Korea for the weights, which is another reason many teams use a hosted id.

A simple test plan

Run the same five prompts on both ids at 768p, five seconds each. The cost is 5 x $0.375 = $1.875 on minimax-h3 and 5 x $0.50 = $2.50 on minimax-h3-max. Compare motion, audio and prompt adherence yourself; this post does not rank their quality, because the pages cited give no Sume-side benchmark.

What the repo table covers

The prices above come from the repo's provider pricing table, multiplied by 1.25 as the Sume docs describe. The H3 family also has h3-max-recast, a separate catalog entry that takes one source video plus one to four person photos at 768p or 1080p for 5 to 30 seconds. It is a different tool from text-to-video, so it is not priced in this comparison.