MiniMax H3 Spicy: The 5-Second Floor and What It Costs

minimax/minimax-h3-{t2v,i2v,ref2v}-spicy cannot generate a clip shorter than 5 seconds. That is not a guideline, it is the `duration` enum: integers `5` through `15`, nothing below and nothing above. Wan starts at 2 seconds and Seedance at 4, so H3 Spicy has the highest duration floor of any adult-adjacent video endpoint Siray carries, and that floor drags a price floor along with it.

Key Takeaways

• duration is a 5-to-15 integer enum. There is no 2-, 3-, or 4-second option and no 30-second option
• The cheapest possible call is $0.40, a 5-second clip at 768p. The base endpoints share the same rate card.
• The family has no seed, no prompt_expansion_enable and no negative_prompt, so a take you liked cannot be reproduced exactly.

The Floor Is in the Enum, Not in the Docs Prose

All three Spicy variants require model, prompt, duration, and size; t2v and ref2v also require aspect_ratio, and i2v requires image. size has exactly two values, 768p and 2k. aspect_ratio offers 21:9, 16:9, 4:3, 1:1, 3:4, 9:16.One thing to ignore: the console Pricing table still renders a priced 4k column. The published spec's size enum has no 4k value, so those rows describe a tier you cannot request. Build against the enum.

What the Floor Costs

Rates below come from the console model-detail Pricing block on 2026-09-17. There is no discount badge on this family. Pricing is linear in duration with no volume break.

That is $0.080/s at 768p and $0.130/s at 2k. The number that matters is the first row: $0.40 is the least you can spend to see anything at all, against $0.09 for a 2-second Wan 3.0 Spicy clip.

Iteration Is the Expensive Part

The family exposes no seed. It also has no prompt_expansion_enable and no negative_prompt. Together those mean a good take is not reproducible: you cannot pin randomness, you cannot steer the prompt rewriting, and you cannot exclude an element by naming it.Combine that with the $0.40 minimum and exploratory work gets expensive fast. Twenty prompt attempts at 768p is $8.00, and none can be re-rolled into a controlled variant. The same twenty 2-second probes on Wan 3.0 Spicy cost $1.80, each with a seed you can hold fixed while you change one word.So the family rewards work where you already know the shot: one sustained 5-to-15 second take where the length is the point. It punishes the short loop you plan to iterate on thirty times. How the 2K tier is priced covers the resolution side of that trade.

The Reference-Image Cliff

ref2v-spicy accepts images, videos and audios, up to 9 of each. The first five images are included; every image after that adds $0.40, independent of resolution.

Nine references cost 5x what five do. Price large reference sets before you build on them — MiniMax H3 ref2v goes deeper on that endpoint, and Seedance 2.5 vs MiniMax H3 compares the wider field.

Developer Notes

Reach for H3 Spicy when the brief is one sustained take, elsewhere when it is a short clip you will run many times. Siray exposes both behind one key, so the choice is a model string, not an integration.Create your free Siray account and price your shot list against both duration floors before you commit.

A Note on Compliance

Siray's NSFW models generate legal, consensual adult content only — CSAM and any illegal content are refused with zero tolerance. MiniMax H3 is a general-purpose video model rather than an adult endpoint, and its Spicy variants are described here in neutral specification terms only: this article makes no claim about what any of them will or will not produce. Moderation across Siray's catalog is not a two-way switch; endpoints sit anywhere from strict to light, and light is not the same as none. MiniMax H3 models are third-party models available through Siray's API, not Siray-hosted models.