China’s Kimi K3 Is Out-And Beating Claude Fable and GPT 5.6 Sol on Key Points

Moonshot AI just released the biggest Chinese version ever released, and it surpassed Claude Fable 5 for the record.

AI Elo writing-the index in which the models write the real texts tested as blind against the published versions, scored using the Elo system with chess players – putting the Kimi K3 at 2,840, above the Fable 5 (max) at 2,760. It is the most ancient Anthropic society.

K3 also took the top spot on Arena AI’s Frontend Code Leaderboard – a ranking made from thousands of peer-to-peer votes on coding performance, as well as an Elo score – with 1,679 against Fable 5’s 1,631. First place in six of the seven front-runners.

The Artificial Analysis Intelligence Index-a score created from nine independent tests covering writing, reasoning, club activities, and knowledge, rated 0 to 100 – ranks K3 at 57, Claude Fable 5 at 60, GPT-5.6 Sol at 59, and Claude Opus 4.8 at 56. 3%.

If you want to have an idea of ​​what he can do, this is it zero results quickly asking the model to create an iOS sketch. In comparison, this it is the best simulation shared on social media using GPT 5.6 Sol and the speed of detail.

What is this thing?

K3 packs 2.8 trillion units – the numbers that store brand information – into a professional mix design. The expert mix divides the fields into 896 “experts” and starts a sub-section for each task. This is how you get the best intelligence without melting the server room.

“It’s the world’s first open source 3-trillion-parameter model, designed for front-end intelligence including long-form text, information, and reasoning,” Moonshot AI. he says. It’s not an advertising show: DeepSeek’s V4-Pro tops 1.6 trillion segments, its Moonshot’s K2 at 1 trillion. K3 doubles the closest heavy open competition on the size chart.

It comes with a window of one million information – symbols are the key part of the AI ​​knowledge, about three thirds of every word – native image and video, and always thinking.

Two methods of construction help to achieve good results. Kimi Delta Attention accelerates rendering of long sequences—up to 6.3x faster for millions of conditions. Residual Information provides information selectively in groups of colors instead of accumulating it uniformly, adding about 25% of training performance under 2% additional calculation cost – together it results in 2.5x better scaling than K2.

Benchmarks are good, Prices are even better

Kimi K3 costs $3 per million input tokens and $15 per million output tokens – the same price as Claude Sonnet 5, Anthropic’s central model. The difference is that Sonnet 5 is the middle part of Anthropic; K3 is three points below Fable 5 on the Artificial Analysis composite. For work on the nine benchmarks, the K3 runs $0.94 versus $1.04 for the GPT-5.6 Sol and $1.80 for the Opus 4.8.

In other words, the brand offers high performance at moderate prices.

Like Decrypt was closed in Maythe price difference between the border of China and America AI ran 15–30x earlier this year. The K3 doesn’t come close to DeepSeek’s price range – it tops out at a mid-range Western model – but it offers near-limit performance there. For teams building on the API this represents a significant change in revenue.

If Anthropic goes ahead with its plans to make Fable 5 available through an API, K3 becomes a more open-source solution than any other brand in the industry—at half the cost of Opus 4.8’s functionality. They are benchmark chasers who are already running the math.

The introduction of K3 is a debate that US chip export regulators do not want to be part of. The US banned Nvidia’s H800 GPUs from being shipped to China until the end of 2023; Moonshot confirmed that it also trained previous models on chips. K3’s benchmark notes the H200s and what the company calls “GPGPUs from another vendor”—commonly interpreted as Huawei Ascend hardware—without specifying where the hardware resides.

Moonshot AI’s president, Yutong Zhang, set the bar straight at Davos this year Silicon Republic: “We knew we didn’t have the chance to just expand… Bank of America researchers, in a statement after the launch, wrote that K3 proves that “expanding education, combined with architectural skills, can still bring the benefits of popular Chinese brands” under these challenges.

The moon shot is one of the so-called Basics of AI Tiger who have changed together the international format without getting chips Washington said they will need. Whether it’s a debate about export controls or an argument that they don’t work, that’s a question Washington hasn’t resolved.

You should read the stars

K3’s AA-Omniscience estimate—an indicator that measures how often the model confidently produces an answer it doesn’t know—jumped from 39% to 51% compared to its predecessor, K2.6. Many correct answers; many remade. The brand also admits in its documentation that it can be “too reactive,” making unexpected decisions on behalf of the user’s long-term autonomy.

For the teams that ran Weapons of Kimi K2.6 and I want to raise, the K3 is a meaningful step in many respects-but that reasonable delta needs to be stress tested before you trust it with anything that needs to be right.

If you want to try it for free, you can. Available on Kimi’s official website. But the good news: The servers are so busy that services are constantly interrupted due to traffic problems, which makes it very easy to use. Another good option is to pay for a subscription or use an API.

The weights will be released on July 27. These will be available for large businesses and businesses. No home GPU, no matter how big it is, can handle this quality.