What a 2K H3 clip actually costs, and why every reseller charges more

Video model pricing is usually a maze of token formulas and per-request minimums. H3 is refreshingly not: you are billed per output second, at a rate that depends only on resolution.

So a clip you can quote from memory:

• 5 s at 768P — $0.40
• 10 s at 768P — $0.80
• 5 s at 2K — $0.65
• 10 s at 2K — $1.30

Prices verified 2026-08-17. This is a live page at MiniMax's end; check it before you build a budget on it.

Two routes to 2K, and they are not the same price

You can render at 2K directly, or render at 768P and regenerate. The second looks cheaper per second and usually is not, because regeneration rebills the source assets.

For an eight-second clip with no reference material:

Identical — which is the point. Regeneration is not a discount route to 2K. It is there so you can review at 768P first and only pay the finishing cost on takes you are keeping. On a shot you nail first time it costs the same; on a shot you iterate five times at 768P it costs a great deal less, because four of those takes never get finished.

Add reference images and the two routes diverge, since regeneration bills the source material again at its own rate — $0.025 per extra image, $0.05/s for input video.

Context-IR is a real line item and it is easy to underestimate

The prompt understanding stage is billed by token, and the documents it produces are large. From the official examples:

Small next to $1.04 of generation. Not small next to a 768P clip you are iterating on twenty times, where it can add a fifth to the bill. Worth knowing it exists before it appears on an invoice unexplained.

Every reseller is above the official price except one

We priced the same 2K second across the market on 2026-08-17:

Two footnotes that matter more than the ordering:

• Segmind's API only accepts 2K. The 768P tier it advertises at $0.1125 is not reachable through the endpoint.
• fal enforces a five-second minimum, so the cheapest possible clip there is $1.30 before you have chosen a length.

Atlas Cloud is worth a second look because it charges one flat $0.14/s regardless of resolution. Against official pricing that is +8% at 2K and +75% at 768P — so it is nearly competitive for finished work and a poor deal for the iteration passes, which is exactly backwards from how most people actually work.

The one provider under official price is selling a different product

PiAPI lists Qubico/minimax-h3 at $0.05/s for 768p and $0.03/s for 512p — genuinely below MiniMax's own rate. That is because it is self-hosted open weights, not the hosted system:

• No 2K. Regenerate-2K has no open implementation.
• No Context-IR, so your prompt reaches the generator as written.
• Three aspect ratios only: 16:9, 9:16, 1:1.
• A 512p tier that does not exist in the official product at all — a reliable tell that you are talking to someone's own deployment.

That can be the right trade for bulk 768p work where you are writing your own structured prompts anyway. It is not the same output, and the price gap is what the two missing stages are worth.

Some other things the pricing page will not tell you

There are no volume packages. MiniMax's video packages page explicitly excludes H3; the Hailuo-era bundles do not apply. Everyone pays list.

The domestic Chinese platform is a different price list. platform.minimaxi.com runs the same model on the same V2 API in yuan: ¥0.80/s at 2K, ¥0.50/s at 768P, ¥0.30/s to regenerate. At roughly 7.2 that is about 13% below the international rates. Same company, same model, two published price lists — the difference is the billing currency. We have not tested the moderation behaviour or rejection rate on those endpoints with English prompts, which is the thing you would want to verify before routing production traffic there.

Output URLs expire. Download or copy results to your own storage promptly, and note that tasks are only queryable for seven days.

Reconcile against usage, not your own maths. A successful task returns total_seconds, input_seconds, output_seconds and input_image_count. Those are the numbers you were billed on.

The short version

For anything that ships, MiniMax direct is the cheapest full-featured route, and the resellers are charging between 7% and 100% for their convenience layer. Iterate at 768P, regenerate the keepers, and quote clients in whole 8-second blocks at $1.04 — it is the only length H3 renders to a whole second anyway.