MiniMax H3 and Seedance 2.5: video generation now competes on price and length

On 31 July MiniMax officially announced H3, the successor to the Hailuo line, designed to handle multimodal context — text, images, video and audio — and to produce video with native stereo audio in a single inference pass. The technical specifications on the official page cap generation at 15 seconds and 2K resolution. On cost, the company states that the per-second price at 2K is less than a third of mainstream models, and less than half at 768p; the absolute price list, however, does not appear in the official documents consulted, which limits the analysis to relative comparisons alone.

A significant part of the announcement concerns distribution strategy: MiniMax said it intends to open the model's weights “in the coming days”, while making the move conditional on the regulatory framework in force. At the time of the announcement the weights were not yet available, which makes the opening a statement of intent rather than a resource anyone can use. According to Reuters, the model targets commercial sectors such as advertising and e-commerce, offering editing of existing content and motion transfer, although the claim about compatibility with Chinese-made chips cannot be independently verified.

On the same day, ByteDance's Seed team confirmed the Seedance 2.5 update, previewed at the Volcano Engine FORCE conference on 23 June. The new model raises the bar it claims for itself, allowing a single generation of up to 30 seconds, with support for multi-step extensions and consistency of scenes and characters. It accepts up to 30 images, 10 videos and 10 audio clips as reference material in a single request, along with timestamp-based local editing. The rollout will follow a staggered schedule, arriving first in Jimeng AI and in the professional version of Doubao, while API access through Volcano Ark is described as coming soon, with no firm date.

Bloomberg noted that the two updates were published a few hours apart. The two launches trace diverging paths in video generation: MiniMax is betting on cheap and potentially open access, while ByteDance consolidates a closed offering built into its own consumer ecosystem. Neither model comes with a technical report or independent benchmarks to support qualitative claims such as less of the “oily” look or consistency across 30 seconds. Without quantitative data that third parties can verify, the quality claims remain the makers' own.

— Olya