analysis

Replicate advises starting with MiniMax H3 Highspeed at $0.06 per second, using H3 for 2K

TL;DR

Replicate recommends starting with minimax/h3-highspeed, dropping to minimax/h3-lite for large batches, and moving up to minimax/h3 for 2K output or difficult shots. A 10-second clip at 768P costs $0.60 on Highspeed, $0.40 on Lite and $0.80 on H3.

3 min read
0

Replicate recommends starting with MiniMax H3 Highspeed, switching to H3 Lite when generating many options, and using H3 when you need 2K output or the shot is hard. In a guide published November 3, 2026, the platform covers three video models in the MiniMax H3 family: minimax/h3-highspeed, minimax/h3-lite and minimax/h3.

Pricing and resolution

Prices are from the Replicate model pages as of October 31, 2026.

Model Resolution Price per output second 10 seconds at 768P
H3 Highspeed 768P $0.06 $0.60
H3 Lite 480P, 768P $0.025 at 480P, $0.04 at 768P $0.40
H3 768P, 2K $0.08 at 768P, $0.13 at 2K $0.80

At 2K, a 10-second H3 clip costs $1.30. Replicate also gives batch figures for a hundred 10-second clips at 768P: $40 on Lite, $60 on Highspeed and $80 on H3.

All three models share the same inputs. They generate clips of 4 to 15 seconds and support text-to-video, image-to-video from a first and optional last frame, and generation guided by image, video or audio references.

Which model for which job

Highspeed is Replicate's pick for iteration. MiniMax pitches it as a balance between quality and speed, and it keeps the full set of reference inputs. It outputs only 768P. Replicate's example, a 5-second 768P text-to-video clip of a hummingbird, took 21 seconds. The guide highlights reference-driven work, such as a fashion spot built from separate images for location, character and product.

Lite is the cheapest option, suited to storyboards, alternate openings and batches where most results will be discarded. Replicate says to pick it to save money, not time: a 5-second Lite clip and a 5-second Highspeed clip both took about 21 seconds on the model-page examples, though the guide notes that one run each proves little and that run time depends on the prompt and load. Replicate also warns that Lite output needs the closest checking, since a clip can look good while swapping the wrong object or dropping a character. For fiddly edits, it suggests testing a few on Highspeed first.

H3 is the top of the family and the only one of the three that outputs 2K. It is also the only one whose resolution defaults to 2K, which matters for cost. Replicate recommends it for motion graphics, choreographed effects and edits where several elements must be right at once. It cautions that 2K does not fix lettering, so text should be checked on every take.

Image-to-video at 2K

Replicate ran seven stills, a mix of photos and illustrations, through H3 using the first_frame_image input. Each clip was 5 seconds at 2K, which costs $0.65. Output keeps the input image's shape: a square still returned 1440 × 1440, a 4:3 still returned 1888 × 1440, and a 16:9 still returned 2560 × 1440.

The seven runs took between two and a half and four and a half minutes. The desert car example, a 5-second 2K clip, took about three minutes. Replicate reports that a painterly tiger illustration stayed painterly and a black and white horse scene stayed black and white without extra prompting. The guide notes that the model-page example videos use different prompts and one starts from an image, so they are not a fair side-by-side comparison.

The guide also includes sample prompts, including a 15-second, 16:9 fashion spot for Highspeed using four reference images in a set order: setting and mood, character, bag, and closing logo.

What this means

The three models differ mainly in resolution and price, not in inputs, so the choice comes down to output needs. Highspeed is limited to 768P, so 2K delivery requires H3. Lite's 480P option carries the lowest per-second rate in the family at $0.025. Replicate's own advice is to verify Lite results closely and to check text on every H3 take, which means review time belongs in any cost estimate alongside the per-second price. Run times in the guide are single-run observations, and the guide itself says they vary with prompt and load.

Related Articles

analysis

OpenAI's Ted Sanders: Pick Astra for quality, Sol for balance, Terra or Luna for price

OpenAI's Ted Sanders says most performance gaps between its models come down to model size, but newer, smaller models can beat older, larger ones. He positions GPT-6.1 Terra against the larger Astra and offers a pick-by-budget rule: Astra, Sol, Terra or Luna.

analysis

Revolut cuts AI coding inference costs roughly 90% by making self-hosted GLM models its default

Revolut says it cut AI coding inference costs by roughly 90% in less than a year by making self-hosted GLM models its default over commercial tools like Claude Code. The fintech serves GLM-5.3 on 56 Nvidia B200 GPUs to roughly 2,000 engineers, according to infrastructure lead David Abitbol.

analysis

OpenAI adds task router to Dots to cut compute as Musk calls Grok Bot costs 'unsustainably high'

OpenAI announced an automatic router that sends Dots tasks to the smaller GPT-6.1 Terra model rather than the larger GPT-6 Astra, to reduce compute use. Elon Musk said this week that Grok Bot's running costs are "unsustainably high relative to revenue." Fortune reports that compute supply is the main constraint on personal AI agents.

analysis

Google Genie 4 points to new uses for world models in robotics and AI testing

Google's updated Genie 4 points to new uses for world models in robotics and AI testing, according to AI Business. The outlet says enterprises must weigh that potential against accuracy gaps.

Comments

Loading...