November 3, 2026

The Chinese model Kimi K3.1 Pro, developed by Moonshot AI, is positioned as a competitor to Claude and GPT in a number of tasks, primarily frontend development. The model also does not face the blocks and restrictions on registering new accounts that, according to the author, Claude does. This article presents the model's stated specifications and the results of a practical comparison with Claude Opus 5.5.

Model Specifications

The developers report improvements in composition, graphics, and animations in the updated version.

The model shows the following results in independent rankings:

• Arena. A platform where users compare the results of different models and choose the best one. Kimi K3.1 Pro ranks eighth among all models in the world and performs on par with Claude Opus 5, at a cost that is 70% lower.
• Artificial Analysis. An overall independent ranking in which the model takes fourth place, three points behind Opus 5.5.
• Factuality stress test. The share of fabricated claims dropped almost threefold and is now lower than Opus's.
• Large-task test. Neural networks generate 20 projects, each of which requires several hours of work. Kimi takes first place on this metric, ahead of Claude Opus.

Additional model capabilities:

• built-in image generation;
• presentation mode;
• 3D work in a single window without additional integrations.

The model is open source. Moonshot promises to make it freely available by November 4. This makes it possible to deploy the model on your own server, provided you have the necessary resources.

Cost

Through the Kimi API, it costs half as much as Opus.

The source text is in Russian, not English, so I'm translating it from Russian into English:

Testing Methodology

Both models received identical prompts and operated at the same reasoning level. Four tasks were evaluated: building a website from scratch, cloning a reference website, developing a browser game, and creating a presentation.

Test 1. Website for an Interior Design Studio

Claude Opus 5.5. Claude cannot generate images, so the author asked it to take photos from the internet. The model created an animation in the main section. When the cursor hovers over the images, they enlarge and the label «View» appears. The website also has:

• an accordion that, in the author's opinion, could have been made smoother;
• a parallax effect;
• a calculator with a slider, where the price changes when the area and housing type (apartment or house) are selected, and a «Discuss estimate» button;
• a standard feedback form;
• a footer whose text fills with color on hover.

The author notes that the website is well built, there are no layout problems, and all the text is readable. At the same time, he noticed signs of neural network involvement: repeating patterns, a scrolling ticker, and a font and numerical figures that, in his view, look artificial.

Kimi K3.1 Pro. The model generated the photos itself. According to the author, the numerical figures look appropriate here thanks to a well-chosen font. Other elements:

• a scroll effect in which the image changes and then its description appears;
• a calculator that, in the author's opinion, looks more impressive, with a «Discuss project» button;
• a footer whose text does not span the full width of the block.

Author's conclusion. The Kimi website is simpler and has fewer animations than Opus 5.5. Signs of neural network involvement are less noticeable in it; however, the author attributes this to being less familiar with Kimi's work.

Test 2. Copying the Reference Website

The reference was the website durable.com, with golden particles, interactive objects, flying coins, a moving map, and a final object. The author recorded the screen and sent the video to both models. No other instructions were given. Kimi spent about three times as long on development as Claude did.

Claude. The model reproduced the falling money effect and the final coin fireworks. The interactivity works, but the motion is somewhat jerky. Scrolling is slow, and objects appear abruptly. According to the author's assessment, the footer was done well.

Kimi. One of the objects appears with a delay and does not respond to interaction. The golden mountain from the reference is missing. The placement of the models, which in the original are located on both sides, has been changed, and they are shown in one place. The block is smoother than Claude's, but in the author's opinion, Claude implemented the structure of the effect more accurately. The model also added a 3D temple object that can be interacted with.

Author's conclusion. Claude completed the task faster and executed the animations better.

Test 3. Browser Game

The task was to create a game about a space robot that flies through the galaxy and the Solar System with realistic planets. The player must approach planets and fly around them.

Claude. The robot is called Rovik. It moves between planets and is controlled with the Shift key for acceleration. The model created moons, Saturn, the sky, and planets. The author notes the following:

• black squares appear on some objects;
• the gas giant cannot be landed on;
• gravity is taken into account: the robot lands on some planets faster than on the Moon;
• the game features black holes and other galaxies;
• one of the planets is based on Earth but with a modified shape;
• the planet Cubis, in the author's opinion, resembles Minecraft.

Kimi. The author rates Earth and space as high quality. The game allows flying to Uranus and Jupiter. The following is noted:

• the robot's animation is slower than in Opus;
• black holes are present;
• a planet similar to Earth but dark;
• spikes are located on one of the planets;
• gravity differs from that of other objects;
• a blue Saturn with water;
• the outer shell of the planets looks better than Claude's.

Author's conclusion. Claude better developed the inner world of each planet and accounted for physics, including the possibility of landing. Kimi wins in the visual aspect of the planets' outer shell.

Test 4. Presentation on the NASA Lunar Program

Both models used photos from the internet.

Claude. The model used orange for accents. The author notes that the combination of black and orange is often used by neural networks. In his assessment, the images are interesting. One planet on a slide looks curved; otherwise, the slides were rated as successful.

Kimi. The main block was rated lower. The model also highlighted accents in orange. The presentations turned out similar. In the author's opinion, one image is less suitable than the one chosen by Claude. The remaining slides were rated as high-quality.

Author's conclusion. There is no clear leader. In some places, Claude selected images for the black slides more accurately.

Итоги

Сильные стороны Claude Opus 5.5:

• более продуманная работа с деталями;
• учет физики и особенностей планет в игре;
• более детальная проработка внутреннего мира объектов.

Сильные стороны Kimi K3.1 Pro:

• визуальная часть сайтов, по оценке автора, выглядит качественно и мало напоминает шаблонный контент, созданный нейросетью;
• автоматическая генерация изображений, которой нет у Claude;
• стоимость API вдвое ниже, чем у Opus;
• отсутствуют, по словам автора, проблемы с регистрацией и блокировками аккаунтов. По его словам, у Claude такие проблемы есть, и большое количество пользователей столкнулось с блокировкой аккаунтов.

Автор делает вывод, что Kimi K3.1 Pro приблизилась к уровню топовых моделей и при этом дешевле в API и доступнее.