Claude Fable 5 just lost its crown. A new model called Kimi-K3, built by China’s Moonshot AI, now writes the best website code, according to Arena’s latest rankings. It scored 1,679 points, while Fable 5 scored 1,631.

Arena ranks AI models simply. People give two hidden models the same task, then vote for the better result. Its web coding board alone counts nearly 470,000 votes across 96 models.

How Kimi-K3 Beat Claude Fable 5

The win came out of nowhere. Moonshot’s previous model, Kimi-K2.6, ranked 18th. Its replacement shot straight to first.

The 48-point lead is no photo finish. Fable 5 sits just 13 points above third-place GPT-5.6 from OpenAI. Kimi-K3 also won six of seven coding categories, covering marketing pages, data dashboards, and consumer apps. Claude Fable 5 kept first place in just one, Gaming.

Anthropic still owns the depth chart, placing nine of the top 20 models. The timing stings anyway. Days ago, Elon Musk called Anthropic the industry leader. Rankings also spark debate, as past AI benchmark disputes over Claude have shown.

Mark Zuckerberg’s Muse Spark 1.1 is also on the list at position 11, only days after roll-out.

Why This Matters Beyond One Leaderboard

Here is the bigger story. Kimi-K3 undercuts its rival on price. Moonshot lists it at $3 per million input tokens and $15 for output. Arena lists Fable 5 at $10 and $50.

Moonshot will also make Kimi-K3’s full model weights available by July 27. Anyone can then run the top coding model for free. That feeds a wider shift, with Chinese models overtaking US rivals in monthly token use.

The rivalry runs both ways. Alibaba told staff to drop Claude Code from July 10 over security fears.

The next few weeks will show whether Kimi-K3 keeps its lead as more votes come in.

Источник: BeInCrypto