Picir.ai

MiniMax H3 Image Pricing: Official API Rates & Cost Guide

on 8 days ago

MiniMax H3 Image Pricing: Official API Rates & Cost Guide

MiniMax H3 Image pricing is based on tokens, with separate rates for text input, reference images, and generated images. The Standard model charges $2.30, $1.50, and $3.80 per million tokens for those categories, respectively. A useful budget also accounts for prompt enhancement and the number of drafts needed before an image is ready to use. Official pricing overview.

Prices checked November 12, 2026. All amounts are USD. The illustrations are conceptual, not MiniMax output samples or measured quality comparisons.

A production estimate needs more than the price of a successful test. The calculations below separate the API charge from the cost of getting an image approved.

MiniMax H3 Image pricing: the official API rates

MiniMax H3 Image API pricing distinguishes MiniMax-H3-Image from MiniMax-H3-Image-Highspeed. Use the exact model name when checking a bill.

Billing categoryStandard, per million tokensHighspeed, per million tokens
Text input$2.30$2.30
Image input$1.50$3.00
Image output$3.80$7.60

The separate H3-Image-Context-IR prompt enhancement endpoint lists $2.30 per million input tokens. These are image-model rates; the H3 video tables use different billing units. MiniMax image pricing.

Highspeed pricing therefore needs its own estimate. Its image input and output rates are twice Standard's, while text input is unchanged. A speed-oriented model name does not imply a cheaper request. Compare the time saved with the extra spend on your actual workload before choosing it for every job.

What does one generated image actually cost?

MiniMax publishes illustrative totals, rather than a guaranteed flat cost per image:

Example taskStandardHighspeed
Text-to-image, 1024 × 1024$0.02179$0.03529
Text-to-image, 2048 × 2048$0.02333$0.03622
Edit with one 1024 × 1024 reference, 2048 × 2048 output$0.02658$0.04576

Actual totals vary with text tokens, including enhancement. Official per-image estimates.

For Standard, calculate the undiscounted request charge as:

Cost = (text input tokens × $2.30 + image input tokens × $1.50 + image output tokens × $3.80) ÷ 1,000,000.

For example, suppose a completed request reports 2,000 text input tokens, no reference-image tokens, and 4,000 output-image tokens. That would cost $0.00460 + $0.01520 = $0.01980. With the same hypothetical usage, Highspeed would cost $0.03500. Those token counts demonstrate the formula; they are not predictions for a particular resolution.

Use the detailed token categories in the returned usage object. The text-to-image API reference describes those fields. Do not substitute prompt word count, image file size, or the total token count multiplied by one rate. Each category has a different price.

Resolution and the temporary 4K discount

Three framed vase images with increasingly fine pixel grids illustrating image resolution

Choose output dimensions explicitly when testing. The image generation guide defines 1K, 2K, and 4K as area levels, with 4K the default. It gives 2720 × 1536 as the 2K, 16:9 output. A “2K” label is therefore not simply a promise that the longest edge is 2,048 pixels.

The current offer halves the total request cost for both models when the delivered image contains at least 16,600,000 pixels. It ends November 17, 2026, at 16:00 UTC, excluding that cutoff. Standard's 4096 × 4096 text-to-image example drops from $0.06879 to $0.03439; Highspeed's drops from $0.13001 to $0.06501. Official offer conditions.

Keep a separate estimate without the discount for work scheduled after that date. A promotional price can make a trial look affordable while understating next month's recurring workload. Record the final dimensions as well as the requested preset so that your estimate can be reconciled with the delivered image.

How prompt enhancement affects image editing costs

Prompt enhancement can make the bill larger than a count of your typed words suggests. It is enabled by default for text-to-image and always runs for image edits. The image editing API says returned text-token usage includes tokens converted from enhancement cost. When calculating from that usage, do not add the same enhancement charge again.

A separate enhancement call is a different matter. The Context-IR documentation explains how to reuse its revised_prompt in text-to-image with disable_prompt_enhanced: true. That avoids paying for enhancement again in the generation step. Passing it to an image-edit request instead triggers enhancement and billing again. The standalone endpoint supports Standard only.

When testing prompt reuse, keep the brief and output dimensions the same. Add the enhancement and generation charges together, then include any additional attempts. Saving money on the last request helps only if total spend falls. Keep related request records together so a reviewer can reproduce the calculation.

Plan a production budget around accepted images

Several vase image drafts beside one selected print and a blank receipt for tracking production costs

Budget for every candidate generated on the way to an approved image. Suppose a project needs 1,000 approved images and you allow four generated candidates for each. Using the official Standard 2K estimate as a planning input, 4,000 × $0.02333 gives $93.32 in generation costs, or $0.09332 per accepted image. This is a hypothetical allowance, not a measured acceptance rate or a quote.

If two candidates prove sufficient, the same calculation falls to $46.66. If eight are needed, it rises to $186.64. This sensitivity matters more than rounding a single-image estimate to the nearest cent. Add editing, storage, and human review costs separately when they apply to your project.

Use a small pilot to replace the assumed candidate count with your own results. Define acceptance before generating: required crop, correct subject, readable product details, and permitted visual changes. Otherwise, a team can keep requesting “one more variation” without learning whether its prompts are improving.

Technical retries need their own control. MiniMax's generation API is not idempotent: a client timeout or disconnect can still leave a successful, billable generation on the server. Retrying starts another generation. The official SDK guidance recommends max_retries=0 and a client timeout of at least 600 seconds. Request and retry behavior.

A practical job record should retain the request ID, model, returned dimensions, usage, and whether the result was accepted. A batch can return HTTP 200 even when some images fail; inspect each item and usage.generated_images. The image editing reference defines billable usage as counting successful images only. Review failed items individually before rerunning a batch. That makes duplicate work easier to identify and keeps technical failures separate from creative rejection.

For a browser-based workflow, start with Picir's AI image workspace and compare its displayed options with your brief. Check Picir's pricing page separately before buying credits. Do not convert the upstream API table into an assumed Picir credit price or assume that every model is available there.

FAQ

Is MiniMax H3 Image billed at a fixed price per image?

No. The published rates use token categories, and the single-image totals are examples. For a useful estimate, record the model, dimensions, reference-image usage, and returned text tokens. For a project budget, also include candidates that generate successfully but are rejected during review.

Is Highspeed cheaper than Standard?

Its listed image-token rates are higher. Whether it lowers your overall production cost depends on the value of saved time and how many outputs meet your requirements. Test both on the same brief and compare spend per accepted result; the model name alone cannot answer that question.

Will disconnecting stop an image generation charge?

No. The API documentation warns that successful generation remains billable after a client disconnect or timeout. A missing response is therefore not proof of a free failed job. Check available request and usage records before resubmitting, and avoid automatic retries that can create duplicate work.

Conclusion

To budget MiniMax H3 Image pricing, record token charges alongside the number of approved images. Start with a small run at your intended dimensions, calculate charges from returned usage, and track how many results you actually accept. Use that measured workflow to plan the larger project.