MiniMax launches M3.1-Flash-Preview: million-token context, native multimodality, replicating Opus 5.5 frontend feats

On September 22, 2025, Anthropic released Claude Opus 5.5, and its frontend generation work was quickly spread widely on X. Six days later, on September 28, MiniMax launched M3.1-Flash-Preview, the latest model in the M series, debuting in MiniMax Code and currently also callable through API subscription plans.

According to Machine Heart, handing Opus 5.5's viral frontend prompt unchanged to M3.1-Flash-Preview not only runs it through, but reproduces it without needing multiple rolls. It should be noted that this reproduction conclusion comes from the experience of the report's author and is a source claim; the report did not provide success rate or stability data.

What is truly noteworthy is not whether a particular demo looks good, but the time gap: Opus 5.5 is a new flagship that was defined just a week ago, and a Flash-tier model can already replicate this part of its capability. This constitutes the core question this article seeks to ask—whether the sinking speed of flagship capability is shifting from being measured in quarters to being measured in days.

The gacha work M3.1-Flash-Preview delivered in the report is titled "ASTRAL SUMMON・Star Spirit Summoning," and the page is a complete wish-making flow: a ten-pull for 1600 diamonds, a SKIP in the top right corner, ten cards flipped open one by one, a special animation when "SSR・NEW" appears, and below the cards a full character panel follows, with level growth, star ascension, and trait unlocks all present.

The second case is the piano-scored hand-drawn animation "mcode and the Black Cat's Piano Piece," in a 1080×1080 frame at 24fps, with the score being an original piano piece of 20 measures in total, with the model completing both the composition and the alignment of the hand-drawn animation.

The report also says M3.1-Flash-Preview can make a retro dance machine game playable and shareable in the browser, use 45 minutes to build a 3D interactive game interface of a spaceship in Earth orbit, and community developers also used it to make a 3D model of "Black Myth: Wukong." These cases are likewise source claims, and the report did not provide verifiable original links and complete processes.

The report also observed a detail: M3.1-Flash-Preview will proactively take screenshots to check its own project during a task, or go back to troubleshoot bugs it previously wrote, only occasionally not quite knowing when to stop.

The explanation given in the report is: frontend code is relatively easy to learn, and what is truly difficult is having "taste," and this first requires building a visual closed loop—the model must be able to "see" its own work.

According to MiniMax's official statement, the M3 series conducts multimodal training from Step 0, and for this purpose rebuilt the pretraining data pipeline and introduced a large amount of interleaved image-text data. The report believes this gives the model flagship-level visual judgment and correction capability. It should be pointed out that this official statement did not provide an original source citation in the source.

The report further says that in the official cases of M3.1-Flash-Preview, converting visual references into interactive code was listed by MiniMax as a separate capability. The report also mentions that on MiniMax's product line there is the H3 video model with full-modal fusion and strong context understanding capability, as well as MiniMax Design combined with an Agent workbench, and that intent understanding and visual style have long been its capabilities in training and productization.

These explanations are attributions from the report and the vendor, rather than independently verified conclusions. Key specifications of M3.1-Flash-Preview such as parameter count and token cost were not given in the source.

The report believes that the way humans and AI collaborate has already gone through several rounds of change: at the earliest it was like filling out a detailed work order, where layout, colors, every piece of text, and button positions all had to be spelled out clearly; later models could catch vaguer descriptions; and with these frontend works as a marker, collaboration has begun to approach "task delegation"—giving intent and direction, and handing over design decisions along with it.

The report infers from this that when delivery is no longer the bottleneck, the value of people in creative work will concentrate on defining problems and aesthetic judgment, while AI's task is to lower the implementation threshold as much as possible. This judgment is an inference of the report, based on the improvement in frontend generation results, rather than quantitative measurement of real workflows.

The report also mentions that those most likely to use it first may be people who already need a large amount of visual content: brands, designers, e-commerce, and game teams. This is likewise an expectation based on the practicality of frontend generation, and the source did not provide adoption data.

The core judgment of the report is: the benchmark capability defined by Opus 5.5 on September 22 was delivered by a Flash-tier model six days later, even faster and at lower cost; the speed of catching up to the frontier model used to be measured in quarters, and now it is measured in days.

This judgment can be inferred from the aforementioned facts, but it is based on a single event, and the sample is clearly insufficient. The report also acknowledges that frontend is only the latest example, and that over the past year long context, tool calling, and coding capability all went through the same sinking—though the source did not give the specific timing and comparison data for these sinkings.

The report also mentions that OpenAI launched GPT 6.1 Sol on September 30, 2025, saying it reflects that "the AI big brother has also started playing the cost-performance game." This matter appears only in this report's retelling, has not been corroborated by other sources, and can only be regarded as a source claim.

The facts that can be confirmed include: M3.1-Flash-Preview launched on September 28, 2025, debuted in MiniMax Code and can be called through API subscription plans, and has a million-token context and native multimodal capability; Claude Opus 5.5 was released on September 22, 2025; the two are six days apart; M3.1-Flash is currently still only a Preview version.

The parts that are source claims include: reproducing Opus 5.5's frontend cases without multiple rolls, specific demo details such as the gacha and piano animation, building a 3D game interface in 45 minutes, community 3D modeling, the model proactively taking screenshots for self-checking, and MiniMax's official statements about Step 0 multimodal training and visual-to-code capability.

The constraints on the evidence are also clear: all evidence comes only from a single Machine Heart report, lacking MiniMax official first-hand materials and third-party independent verification; the report did not explain the direct comparison test method between M3.1-Flash-Preview and Opus 5.5 under equivalent conditions; nor did it provide pricing, open regions, or a formal version timetable. In addition, the capability of the Preview version may differ from the formal version.