What Does MiniMax Actually Cost in 2026?
MiniMax-M3 and MiniMax-M2.7 are both reported at $0.30 per million input tokens and $1.20 per million output tokens on MiniMax pay-as-you-go pricing, a rate described as a "permanent 50% off" against a $0.60 and $2.40 reference price. Those figures come from third-party coverage and platform documentation rather than a first-hand read of platform.minimax.io, so treat every number on this page as reported rather than confirmed.
That caveat matters more than the headline, because most launch-week coverage repeated the $0.30 figure without it. The second thing coverage skipped is the tier line: the reported rate is described as applying to standard requests up to 512,000 input tokens, while M3 is reported to advertise a 1M-token context window.
MiniMax is also one of the few names in this category with no startup credit program AI Perks could find. What it has instead is open weights, which changes the question from "how do I get MiniMax credits" to "who pays for the GPUs." AI Perks tracks $7.7M in credits across 194 companies for exactly that second question.
MiniMax API Pricing, and How Solid Each Number Is
The figures below circulate through third-party coverage and platform documentation. None was fetched first-hand from MiniMax's own pricing page for this article. Prices are per million tokens, in US dollars.
| Model and tier | Input | Output | Cache read | How solid |
|---|---|---|---|---|
| MiniMax-M3, standard, up to 512K input | $0.30 | $1.20 | $0.06 | Reported, several sources agree |
| MiniMax-M3, undiscounted reference | $0.60 | $2.40 | not reported | Reported as the pre-discount list |
| MiniMax-M2.7 | $0.30 | $1.20 | not reported | Reported, conflicting |
| MiniMax-M2.7 via OpenRouter | $0.21 | $0.84 | not reported | Reseller rate, not vendor pricing |
Two things in that table deserve attention.
The first is the discount framing. $0.30 and $1.20 are presented as a "permanent 50% off" against $0.60 and $2.40. A permanent discount is a list price with extra steps, but it is also a rate a vendor can withdraw without ever announcing a price increase. Budget on the reported $0.30 and keep the reported $0.60 in mind as the ceiling.
The second is the M2.7 conflict. One source reports $0.30 and $1.20, matching M3. OpenRouter lists $0.21 and $0.84 for the same model, but OpenRouter is a reseller setting its own margin, so it is not evidence of MiniMax's own price, and the platform.minimax.io figure could not be confirmed either way. If M2.7 economics decide your architecture, price it from your own console rather than from any page, including this one.
Cache reads are the one place the reported numbers are unambiguous: $0.06 per million against $0.30 for fresh input, an 80% saving on anything you send more than once. That ratio, if it holds, is a bigger lever than the choice between M3 and M2.7.
Why the 512K Line Matters More Than the Headline Price
The reported $0.30 rate is described as applying to standard requests up to 512,000 input tokens, while MiniMax-M3 is reported to carry a 1M-token context window. What a request costs above that line has not been published anywhere that could be verified.
That gap is the single largest unknown in MiniMax pricing. The reference price behind the discount is $0.60 and $2.40, exactly double, which is why so much coverage assumes long requests simply bill at twice the rate. Assume is the operative word: no worked example, no second rate card, no vendor statement. Nobody outside MiniMax knows whether a 900K-token request doubles in full, doubles only the portion above the line, or is declined.
| Request shape | At the reported $0.30 and $1.20 | If long requests bill at double |
|---|---|---|
| 200K in, 5K out | about $0.07 | same, under the line |
| 500K in, 10K out | about $0.16 | same, under the line |
| 900K in, 10K out | about $0.28 | about $0.56 |
The spread between those last two columns is the whole uncertainty, and it is a 2x swing on your largest requests. Until the rate is published, plan against the higher column and treat the lower one as upside. Trimming one retrieval step to land a pipeline under 512K input removes the question altogether, which makes it worth more than any prompt tuning you will do this quarter.
What MiniMax H3 Costs, and What the Vendor Has Not Published
MiniMax H3, the model behind Hailuo 3.0, is reported to have launched on July 31, 2026, with open weights following on Hugging Face on August 3 under the MiniMax H3 Community License. It is described as generating 4 to 15 second multi-shot video at up to 2K, with native stereo audio carrying dialogue, effects and ambience, produced in a single pass. No official per-second price for it could be located.
That last sentence is the one to keep. Every absolute per-second figure for H3 in circulation traces back to API aggregators and vendor-adjacent pages rather than to a first-party MiniMax pricing page.
| Reported H3 figure | Amount | Source quality |
|---|---|---|
| Per second at 768p | $0.08 | Aggregator, unconfirmed |
| Per second at 2K | $0.13 | Aggregator, unconfirmed |
| Per second, H3 Max variant | $0.05 to $0.08 | Aggregator, unconfirmed |
| Reference images | $0.04 each, first five free | Aggregator, unconfirmed |
A separate third-party aggregation puts 2K generation in a $0.073 to $0.120 per second band, which brackets that $0.13 figure from below rather than agreeing with it. Several sources landing near each other is weak evidence at the best of times, and here they do not even fully agree.
If the $0.13 figure holds, a ten second 2K clip with synced stereo audio lands near $1.30, and the same clip at 768p near $0.80. Treat those as planning estimates with a wide error bar, not quotes. On the consumer side, the Hailuo app runs a credit-based subscription rather than per-second billing, reported at 15 to 80 credits per video, and no separate API free tier for H3 has been confirmed. AI Perks tracks the video credit programs that offset that spend.
Where the Free MiniMax Credits Actually Are
No MiniMax startup credit program was found, and no free API tier is confirmed in vendor documentation. The widely repeated ¥15 trial credit for new accounts appears in third-party guides only and could not be verified against MiniMax's own docs.
¥15 converts to roughly two US dollars at recent exchange rates. If both that figure and the reported $0.30 input rate hold, it buys a few million input tokens: an evaluation budget rather than runway, and an amount that reportedly moves with promotions.
The real free lever sits elsewhere. M3 weights are reported to have reached Hugging Face within a week of the model's June 2026 announcement, and H3 weights followed in August under a community license, so the models themselves cost nothing to obtain. What they cost is GPU time.
That reframes the search completely. You are not hunting for MiniMax credits, you are hunting for compute credits to run MiniMax weights, and those exist in volume:
| Where the money is | What it covers |
|---|---|
| Cloud and GPU programs | The hardware that runs open weights |
| Inference platform credits | Hosted endpoints, no GPUs of your own |
| Video and media credits | Generation spend, including Hailuo-class output |
AI Perks covers all three categories across 194 companies, with current terms rather than launch-day figures.
How to Run MiniMax Without Paying List Price
Three levers matter here, in rough order of size.
Compute credits beat model credits. For an open-weights vendor, the bill you can actually erase is the GPU bill, not the token bill.
Cache aggressively. Cache reads are reported at $0.06 per million against $0.30 for fresh input, the largest single saving on offer.
Price the self-host option before committing. Weights are free under the community licenses, so the real comparison is your GPU bill against a reported $0.30 per million. With credits from AI Perks, that GPU bill can be zero for a good while.
Frequently Asked Questions
How much does the MiniMax API cost?
MiniMax-M3 and MiniMax-M2.7 are both reported at $0.30 per million input tokens and $1.20 per million output on MiniMax pay-as-you-go pricing, described as a permanent 50% discount against a $0.60 and $2.40 reference price. Those figures come from third-party coverage rather than a first-hand read of the vendor page, and what requests above 512K input tokens cost has not been published. Compute credits that offset this are tracked at getaiperks.com.
Does MiniMax give free credits to new accounts?
No free API tier is confirmed in MiniMax's own documentation. Third-party guides report a ¥15 trial credit for new accounts, roughly two US dollars at recent exchange rates, but the vendor has not published that figure and it could not be verified. No MiniMax startup credit program was found either. AI Perks tracks the compute credits that do exist.
Is MiniMax-M3 open source?
MiniMax-M3 publishes open weights on Hugging Face under a MiniMax community license. Reported specifications put it at roughly 428B total parameters with about 23B active, a 1M-token context window, and native text, image and video input, none of which was independently confirmed for this page. Open weights are not the same as open source, and the license carries its own conditions worth reading before commercial use.
What does MiniMax H3 cost per second?
No official per-second price for H3 could be located. Third-party aggregators list $0.08 per second at 768p and $0.13 at 2K, while a separate aggregation puts 2K in a $0.073 to $0.120 band. All of those are unconfirmed by the vendor, and they do not fully agree with each other.
Is MiniMax cheaper than the large US labs?
Not answerable from public numbers with any confidence. Any real comparison means quoting a competitor's current rate, and those move often enough that a published figure ages badly. The durable differentiators are the reported 1M context window and the open weights, not a headline price that is itself framed as a promotion.
Can I self-host MiniMax models for free?
Yes. M3 and H3 weights are published on Hugging Face under MiniMax community licenses, so the models themselves cost nothing. Your cost becomes GPU time, which is exactly what cloud and compute credit programs cover. AI Perks tracks $7.7M across 194 companies.
The weights are free. Make the GPUs free too.