Accounts, Pricing, and the Cost Model
- →Locate the right key type in the console and authenticate with a Bearer header
- →Price any generation job before submitting it: output seconds plus input material
- →Plan throughput around concurrent-task limits (CONN), not requests per minute
Two kinds of keys, one kind of auth
Everything starts at the console: log in at platform.minimax.io/login; the account area lives at platform.minimax.io/user-center/basic-information. Two different credentials are issued in two different places, and confusing them is the first trap:
- Pay-as-you-go API Key — created under Account > API Keys. This is the key this course uses for every H3 call.
- Subscription Key — issued under Billing > Token Plan for Token Plan subscribers. Per the official FAQ, it is a different key type and is not usable for direct pay-as-you-go H3 calls.
Authentication itself is refreshingly simple. Every H3 endpoint takes a single header:
Authorization: Bearer <token>
That Bearer header is the entire auth story. The classic MiniMax v1 APIs required a GroupId; the H3 v2 API does not — GroupId appears nowhere in any H3 endpoint's documentation. If a guide tells you to append a GroupId to your H3 calls, it was written against the old API. Also send Content-Type: application/json explicitly on every creation request.
What generation costs
H3 bills output by the second, at a rate set by resolution:
| Output | Rate |
|---|---|
| MiniMax-H3 at 2K | $0.13/second |
| MiniMax-H3 at 768P | $0.08/second |
Input material is billed separately:
| Input material | Billing |
|---|---|
| Audio | Free |
| Images | First 5 free; $0.04 per additional image |
| Video | Input video duration x the output resolution rate (2K $0.13/s, 768P $0.08/s) |
Two more line items complete the price list. Video Regeneration — the 768P-to-2K mastering step you will meet properly in the core track — bills at $0.05 per second of regenerated output, with its own input-material schedule: audio free, first 5 images free then $0.025 per additional image, and input video billed at $0.05/s based on the original 768P task's input duration. H3-Context-IR, the hosted prompt-enhancement stage, bills like an LLM: $0.90 per million input tokens, $3.60 per million output tokens.
Worked examples
A one-shot 12-second 2K clip from text. 12 x $0.13 = $1.56. No input material, nothing else to add.
An 8-second 768P reference generation with 7 reference images and a reference audio clip. Output: 8 x $0.08 = $0.64. Images: the first 5 are free, so 2 paid at $0.04 = $0.08. Audio: free. Total $0.72.
The draft-then-master pattern. Generate 10 seconds at 768P (10 x $0.08 = $0.80), review it, then regenerate at 2K (10 x $0.05 = $0.50). Total $1.30 — exactly what 10 seconds of direct 2K costs (10 x $0.13). The arithmetic is not a coincidence you can exploit for savings; the value of the two-step path is the checkpoint: you review the cheap 768P draft before committing to the master, and a rejected draft only cost you $0.80 instead of $1.30. When you expect to iterate, iterate at $0.08/s and master once.
Running out of balance is a first-class, documented failure mode: HTTP 402, insufficient_balance_error. If your pipeline treats every non-200 as "retry later," an empty wallet looks like an outage. Budget before you batch.
The rate limit is concurrency, not RPM
Most write-ups mention rate limits generically, as if H3 had a requests-per-minute budget like other APIs. It does not. MiniMax-H3 Video Generation V2 is limited by CONN — maximum concurrent tasks: 2 concurrent tasks on the free tier, 15 on the paid tier. This is a materially different mental model, and it is unlike MiniMax's own LLM APIs, which are RPM-based.
Practical consequences: submitting requests faster does not help and does not hurt — what matters is how many tasks are in flight (queued or running) at once. A batch job on the paid tier should hold at most 15 open tasks and submit a new one as each finishes, like a semaphore. On the free tier, that semaphore has 2 permits. The production-operations lesson in the core track builds this into a real pipeline.
Token Plan subscribers: read the exclusion
MiniMax sells Token Plan subscriptions at $20, $50, and $120 per month. If you have one, do not assume it covers H3.
The Token Plan docs are explicit: "A small number of special models (MiniMax H3, voice design, rapid voice cloning, etc.) are not currently supported [by Token Plan quota]. Credits usage is unrestricted." Your H3 usage bills from Credits/pay-as-you-go even while your subscription covers other models — and the Subscription Key itself does not work for direct pay-as-you-go H3 calls. For this course, use a pay-as-you-go API key.
One genuinely free on-ramp exists: MiniMax Hub offers "3 times for free" H3 trial generations, outside the metered API. That is a fine way to see output quality before creating a key, but it is not an API allowance — the metered API's only free elements are the CONN=2 free-tier concurrency and the first-5-images input allowance.
What to take into the next lesson
You now have a key, the Bearer header, and a cost model you can run in your head: output seconds times the resolution rate, plus paid images past the first five. In h3-f3 you will spend your first ~$0.65: one real 5-second 2K generation, submitted, polled, and downloaded through the actual API — including the status enum that a surprising number of guides get wrong.
Set up your account and price three jobs on paper
0 / 4 stepsYou have a working pay-as-you-go key, and your worksheet reads: A = $1.56 (12 x $0.13); B = $0.72 (8 x $0.08 output + 2 paid images x $0.04, audio free); C = $1.30 ($0.80 draft + $0.50 regeneration) — exactly the same dollars as direct 2K, but with a reviewable 768P checkpoint in the middle.
Knowledge check
3 questionsPick an answer to see why it is right or wrong — including the wrong ones.
How is MiniMax-H3 video generation rate-limited?
What does a 5-second 2K generation with no input material cost?
You pay $50/month for a MiniMax Token Plan subscription. How is your H3 usage billed?