"Impressive": Elon Musk surrenders to Kimi K3, the Chinese open model that cannot cope

It's good, it's pretty and it's (quite) cheap. Kimi K3, the new open AI model from Chinese startup Moonshot AI, is causing such a sensation that the company has had to pause new subscriptions because it can't keep up with the demand.

Kimi K3 is the largest open weights AI model ever published, with numbers that probably rival those of the frontier models from Anthropic and OpenAI, which do not provide information on the size of their models. Those 2.8 billion parameters make the difference and are a good part of the reason why it represents a real leap in quality according to all published benchmarks.

Elon Musk himself published a brief "Impressive" on his X account in response to the very complete performance analysis of Artificial Analysis. Its agentic behavior surpasses that of Opus 4.8 and only Fable 5 surpasses it; in a specific benchmark it goes even further and is the best of all the models evaluated, including those from OpenAI and Anthropic.

In programming it is better than Opus 4.8 and GPT-5.5, but inferior to Fable 5 or GPT-5.6. No Chinese model had come so close: GLM-5.2, although notable, competed more with GPT-5.5 and Sonnet 5 than with the US frontier models.

DeepSeek demonstrated that it was possible to access truly capable models at a very affordable price, and GLM-5.2 proposed the same: achieving 90% of the capacity of frontier models at 20% of the cost. With Kimi K3 the trend changes: it is more affordable than Fable 5 or GPT-5.6, but not as much as one would expect. The cost per million input/output tokens is $3/15, while Fable 5 costs 10/50, Opus 4.8 costs 5/25, and GPT-5.6 Sol costs 5/30.

One of the influencing factors is the large number of tokens that Kimi K3 uses when responding. It is a model that "thinks a lot", which improves precision but increases the bill: the cost per task in the Artificial Analysis test battery is 0.95 dollars, at the level of 1.04 for GPT-5.6 Sol, cheaper than Fable 5 (2.75) but much more expensive than Grok 4.5 (0.31) or GLM-5.2 (0.47).

Analyst Simon Willison tested the model with his usual test: generating an SVG image of a pelican on a bicycle. The result was of very good quality, but it consumed almost 17,000 tokens at a cost of 25 cents. It is not conclusive, but it reveals that for a simple task the result, although outstanding, is not particularly efficient.

Unlike Anthropic or OpenAI, Moonshot AI does not mention cybersecurity capabilities in the release notes. Even so, Vercel CTO, Malte Ubl, explained that after several tests he found it to be a very useful model for finding and correcting vulnerabilities.

The expectation has been such that the company has paused new subscriptions in order to respond to all requests without harming the experience of current users. A striking decision that makes a reality clear: they cannot cope.