NotCheap MAKE EVERY CREDIT COUNT
Analyze Your Costs

MiniMax API Pricing: Text, Speech and H3 Video

Share this guide

MiniMax bills text, speech, and video with three different meters: tokens for text reasoning, character counts for speech synthesis, and rendered seconds for video generation. The current developer lineup centers on M3 text models, Speech 2.8 audio, and the flagship H3 video engine. To build a budget you can trust, model your volume in each category separately. And don't assume a consumer Hailuo app membership or a third-party coding subscription covers your developer API usage, because it doesn't.

All pricing figures were verified 25 October 2026 against the official pricing overview, quoted in USD before sales tax. The production workloads and regeneration assumptions below reflect our standard engineering test scenarios.

Three different meters

ServiceSelected model or tierListed unit price
TextM3.1-Flash$0.15 input, $0.015 cache-read, $0.60 output per million tokens
TextM3, standard, up to 512K input tokens$0.30 input, $0.06 cache-read, $1.20 output per million tokens
TextM3, standard, above 512K input tokens$0.60 input, $0.12 cache-read, $2.40 output per million tokens
Speechspeech-2.8-turbo$60 per million characters
Speechspeech-2.8-hd$100 per million characters
VideoH3 at 768p$0.08 per output second
VideoH3 at 2K$0.13 per output second
VideoH3-Highspeed at 768p$0.06 per output second
VideoH3-Lite at 768p$0.04 per output second

Think of this overview as a rate card, not a promise that these models deliver interchangeable quality. Two things to keep in mind: M3 Priority carries an extra routing surcharge, and H3 input video is billed separately by duration. The older Hailuo 2.3 model family also still shows up under legacy endpoints. Those historical per-clip rates are easy to mistake for current H3 pricing, so don't let them fool you.

A coherent brief: 100 six-second narrated clips

Let's walk through a practical multimedia campaign: producing 100 approved social media video clips, each exactly six seconds long. At a standard narrative pace of 900 speech characters per minute, each six-second script comes to roughly 90 characters. Add a realistic 1.2 speech regeneration factor for tone and pronunciation tweaks, and the full voiceover needs 10,800 synthesized characters. It's worth checking your script cadence and punctuation pauses during pilot recordings, since delivery speed directly drives character volume.

On the video side, assume an average of two candidate generations per final approved clip. That's 200 generation requests and 1,200 billed output seconds. We pick H3 768p text-to-video with at most one reference image per call. Under current flagship H3 rules, the platform includes the first five input images per task at no extra charge, so this single-image workflow has zero reference-image surcharges.

For storyboard ideation and script generation, we budget aggregate totals of 1.0 million uncached input tokens, 2.0 million cache-read tokens, and 200,000 output tokens. In our setup, every M3 request stays well within the standard up-to-512K input tier.

ComponentCalculationUsage charge
M3 uncached input1M × $0.30$0.30
M3 cache reads2M × $0.06$0.12
M3 output0.2M × $1.20$0.24
Speech 2.8 Turbo10,800 ÷ 1M × $60$0.648
H3 768p output1,200 seconds × $0.08$96.00
Total, rounded after adding$0.30 + $0.12 + $0.24 + $0.648 + $96$97.31

Text on its own costs a mere $0.66. If those two million cache-read tokens were processed as cold, uncached input instead, text charges would rise to $1.14. Before you bake cache savings into a forecast, confirm your actual cache hit rates in your developer telemetry rather than assuming them.

Scale the campaign up to 1,000 approved clips at exactly ten times the workload, and the unrounded consumption cost is $973.08. If you multiply the pre-rounded $97.31 by ten, you'll end up two cents off. It's a tiny gap, but the habit is worth keeping: carry full precision through your model and round only at the invoice stage.

What changes the bill most?

At this scale, video generation is the vast majority of the bill. Dropping to H3-Lite at 768p ($0.04/s) cuts the video line in half, from $96.00 to $48.00, and brings total combined usage down to $49.31. The practical catch is that H3-Lite uses a streamlined architecture. Test visual fidelity and temporal consistency before you assume those savings come free.

If your creative bar calls for three candidate attempts per approved clip instead of two, video output grows to 1,800 seconds ($144.00). That lifts total project spend to $145.31 with text and voice held constant. Better prompt engineering that avoids messy retakes will save you far more than sweating fractions of a cent on text tokens.

Video-to-video references add another cost layer. MiniMax bills input video by duration at the chosen resolution. Feeding six seconds of reference video into each of the 200 flagship 768p requests adds 1,200 × $0.08 = $96.00 in input charges, which pushes the project total to $193.31. That input fee applies whether or not the output matches the input duration.

One more thing to keep in mind: H3-Max has different input parameters than the flagship model, including different rules for free reference images. Allowances from one model tier don't automatically carry over to another.

Speech, cloned voices and package access

Upgrading to Speech 2.8 HD ($100/M characters) for that same 10,800-character script costs $1.08, compared to $0.648 on Turbo. HD isn't automatically the right call. Benchmark both models on your actual speaker voice and technical vocabulary before committing to the higher rate.

Custom voices come with one-time setup fees. MiniMax currently lists voice design at $3.00 and rapid voice cloning at $1.50, billed when the voice is first deployed for synthesis. Budget your trial generation characters separately, and make sure you hold verified commercial rights to any voice you clone.

Bundled credits deserve a little extra care. MiniMax sells audio subscriptions, video packages, and pay-as-you-go developer balances under separate policies. The official documentation explicitly notes that promotional video packages support legacy Hailuo models but not H3, so an active package balance won't pay for flagship H3 video calls. Check model compatibility rules before buying any discounted bundle.

Cash paid is not the same as usage consumed

Our $97.31 baseline reflects raw metered consumption before tax. It is not a mandated prepayment tier or a recurring subscription fee. If your account already holds prepaid balance, your out-of-pocket cash for this project may be $0.00 while credits burn down. If you deposit $120.00 as an operational top-up and burn the unrounded $97.308, approximately $22.69 remains for later workloads.

Keep purchased cash balances separate from promotional grants, and watch expiration windows and refund restrictions. Setting wallet threshold alerts in the developer console is a cheap way to avoid surprises. Consumer Hailuo app subscriptions, specialized M Plan allowances, and standard API wallet funds shouldn't be treated as shared cash unless they're explicitly linked in your account settings.

Our recommendation: split your media pipeline by task. Use M3 models for text ideation, Speech 2.8 Turbo for verified narration, and H3 768p for approved video scenes. For long-form audio narration, take a look at our podcast budget. For broader video generation benchmarks, see our video cost comparison.

PASS IT ON

Useful for your team?

Share this guide with a colleague.