Short answer. Stay on MiniMax when you need M3 text and H3 video from one vendor, especially at low or bursty volume. Choose Plugsky for a flat bill, 39 documented model IDs and private deployment. Before you compare the two, note that
plugsky-minimaxis a legacy compatibility ID served by NVIDIA Nemotron 3 Nano Omni 30B, not MiniMax M3 and not the Ultra 550B pool.
MiniMax publishes separate rate cards for text, video, image and audio. Text has two input bands at 512K tokens and a 1.5× priority tier. Video prices output seconds and bills input video and extra images. Audio is per hour for transcription and per million characters for synthesis. Mixing those units is where comparisons go wrong.
How we evaluated this
We weighted five criteria for a team buying a multimodal API, not a single text model.
- Unit fit, 30%. Whether the vendor prices the thing you ship: tokens, output seconds, characters or images.
- Text economics, 25%. Input and output rates at a stated mix, including the 512K band.
- Video and audio capability, 20%. Resolution, duration, input billing and voice options.
- Billing predictability, 15%. Pay-as-you-go versus a flat plan, and how many line items a job creates.
- Deployment control, 10%. Hosted only, private VPC, on-premises or air-gapped.
We read MiniMax's official pricing page on 20 November 2026. We did not generate video or audio, run a quality evaluation or test latency. No 2027 benchmark is implied.
What MiniMax charges, by unit
| Product | Unit | Price | Note |
|---|---|---|---|
| MiniMax-M3, ≤512K input | 1M tokens | $0.30 in / $1.20 out | Permanent 50% off $0.60 / $2.40 |
| MiniMax-M3, >512K input | 1M tokens | $0.60 in / $2.40 out | Same model, higher input band |
| MiniMax-M3.1-Flash | 1M tokens | $0.15 in / $0.60 out | Cheaper fast tier |
| H3, 2K | Output second | $0.13 | Input video also $0.13/sec |
| H3, 768P | Output second | $0.08 | First 5 images free |
| H3-Lite, 480P | Output second | $0.025 | Draft and batch work |
| Speech recognition | Hour | $0.38 | Streaming and diarization |
| Speech synthesis | 1M characters | $60 turbo / $100 HD | Voice design $3 per voice |
| plugsky-minimax | Flat plan | From $8/month | Legacy ID; served by Nano Omni 30B, not MiniMax |
Text economics: the 512K boundary matters
M3 is priced in two bands. Below 512,000 input tokens, the current rate is $0.30 input and $1.20 output per million; above it, $0.60 and $2.40. At a 70/30 input-to-output mix, that is about $0.57 per million below the boundary and $1.14 above it. M3.1 Flash is $0.285 at the same mix.
A single long prompt can cross the boundary and reprice the whole request, so the boundary is an architectural constraint, not a footnote. If your product sends 400K-token documents, you are in the cheap band; if it occasionally sends 600K, those requests cost twice as much. Measure the input distribution before you forecast.
The naming trap: plugsky-minimax is not MiniMax
Do not assume the ID means the weights. plugsky-minimax is a legacy compatibility ID served by NVIDIA Nemotron 3 Nano Omni 30B (A3B), with a 262,144-token context and up to 8,192 output tokens. It is not MiniMax M3, and it is not served by Nemotron 3 Ultra 550B. The same Nano 30B engine serves plugsky-gpt-oss, plugsky-deepseek-flash and plugsky-coder-fast. The live API's served_by field and the x-plugsky-served-model response header are authoritative.
That correction matters for evaluation. If your test depends on MiniMax's tokenizer, long-context behaviour or video style, test MiniMax itself. If you only need a fast text model behind an OpenAI-compatible API, the legacy ID may still fit, but judge it on the engine that actually serves it. The model catalogue lists each ID and its engine.
Video and audio: price the finished asset
Video is priced per output second, so a six-second clip is the useful unit. H3 2K is $0.78; 768P is $0.48; H3-Lite 480P is $0.15. Input video is billed at the output resolution, the first five images are free, and extra images are $0.04 each on H3 or $0.02 on H3-Lite. Regeneration bills the original input again, which is easy to miss in a retry loop.
Audio is unit-specific. Transcription is $0.38 per hour. Synthesis is $60 per million characters for turbo or $100 for HD. A 1,000-word script is roughly 6,000 characters, so about $0.36 to $0.60 before voice design or cloning. A narrated video carries token, video and character costs; add them before comparing vendors.
Cost math at two volumes
Assume a 70/30 input-to-output mix, standard rates, no cache hits and no priority tier. For text, compare M3, M3.1 Flash and Plugsky Builder.
10M tokens:
M3 (<=512K): 7M x $0.30 + 3M x $1.20 = $2.10 + $3.60 = $5.70
M3.1 Flash: 7M x $0.15 + 3M x $0.60 = $1.05 + $1.80 = $2.85
Plugsky Builder: $60
100M tokens:
M3: $57.00
M3.1 Flash: $28.50
Plugsky Builder: $60.00
100 six-second H3 2K clips:
600 seconds x $0.13 = $78.00
100 six-second H3-Lite 480P clips:
600 seconds x $0.025 = $15.00
At 10 million tokens, per-token is far cheaper than any flat plan. At 100 million, M3 is still cheaper than Builder; M3.1 Flash is about half. The flat plan wins when the mix includes enough frontier tokens to cross its break-even, or when you value one bill and private deployment above the per-unit price. It does not win on cheap text alone.
When MiniMax is the better choice
MiniMax wins when the product needs native video and audio beside text, when a single prompt genuinely exceeds 512K tokens, or when per-second video billing matches how you ship. It also wins at low or bursty volume, where a subscription would sit idle, and when the team wants one vendor for M3 text, H3 video and speech.
When Plugsky is the better choice
Plugsky wins when you want a flat monthly bill, one OpenAI-compatible API and 39 documented IDs across chat, coding, vision, translation and embeddings. Builder and Scale have no monthly token cap and no per-token overage; fair-use concurrency and RPM apply. Private VPC, on-premises and air-gapped deployment are available by contract. The trade is explicit: the public 39-ID catalogue does not include MiniMax weights, and its video endpoint should be quoted per job rather than assumed to match H3's per-second rate. See the fair-use definition.
What we did not verify
We did not verify video quality, effective resolution, generation time, retry behaviour or audio quality. We did not test the 512K boundary on real prompts or confirm how MiniMax counts image and video input tokens. We did not verify rate limits, enterprise terms, regional processing or whether the 50% text promotion will remain permanent despite the page saying so. We did not verify Plugsky's per-job video or image credit pricing.
For Plugsky, we did not verify a SOC 2 certification. Request the current SOC 2 report or attestation under NDA and verify the period. Model mappings were checked against the 20 November 2026 ladder and should be re-checked against the live API before publishing.
Frequently asked questions
Is plugsky-minimax the same as MiniMax M3?
No. plugsky-minimax is a legacy compatibility ID served by NVIDIA Nemotron 3 Nano Omni 30B, with a 262,144-token context. It is not MiniMax M3 and not the Ultra 550B pool. plugsky-gpt-oss, plugsky-deepseek-flash and plugsky-coder-fast also run on Nano 30B.
How much does MiniMax M3 cost per million tokens?
On the current permanent 50% promotion, M3 is $0.30 input and $1.20 output up to 512K input tokens; above 512K it is $0.60 and $2.40. At a 70/30 mix that is about $0.57 or $1.14 per million.
How much does MiniMax video cost per second?
H3 ranges from $0.025 to $0.13 per output second. A six-second clip is about $0.15 at H3-Lite 480P and $0.78 at H3 2K, before input video or extra images.
Can I migrate text from MiniMax without rewriting code?
For OpenAI-format text calls, usually yes: change the base URL, key and model ID, then re-test streaming, JSON mode, tool calling and the 512K input boundary. Video and audio need endpoint-level testing for submission, polling and billing units.
Is MiniMax cheaper than a flat-rate API?
At low volume and on cheap text models, usually yes. At 100 million M3 tokens, about $57 is still below Plugsky Builder at $60, while M3.1 Flash at $28.50 is well below it. A flat plan wins on frontier-heavy or multimodal mixes, not on cheap text alone.
Does Plugsky serve MiniMax video models?
Plugsky has image and video endpoints, but the public 39-ID catalogue does not include MiniMax weights, and the legacy plugsky-minimax ID is a text model on Nano 30B. Quote the video job separately and compare cost per finished asset.
Sources
- MiniMax pricing overview: text, video, image and audio — checked 20 November 2026.
- MiniMax documentation index — checked 20 November 2026.
- MiniMax API reference — checked 20 November 2026.
Last checked 20 November 2026. Promotions, input bands and video rates change; confirm the current rate card before committing. Not paid.