Kimi-K3 Hits 1,679 Points and Undercuts Claude Fable 5 by 70% on Price

China’s Moonshot AI just took the top spot in Arena’s web coding rankings. Kimi-K3 scored 1,679 points, beating Claude Fable 5, which landed at 1,631. It’s a gap that looks modest on paper but carries real weight in a field where fractions of a point separate models that cost enterprises millions to deploy.

The jump is pretty stunning when you look at where Moonshot was sitting before. Kimi-K2.6, the previous model, held the 18th position. Kimi-K3 didn’t just climb — it basically leapfrogged 17 competitors in one move. That kind of vertical rise doesn’t happen often in these rankings, and it’s got a lot of people in the AI community paying close attention. The model took the top slot in six out of seven coding categories tracked by Arena, including marketing pages, data dashboards, and consumer apps. Claude Fable 5 held on in exactly one category: gaming. That’s it.

One category. Out of seven.

What Kimi-K3 Actually Beat

Anthropic isn’t going anywhere. It still owns nine of the top 20 models in Arena’s rankings — a kind of quiet dominance that’s easy to overlook when a flashy newcomer grabs first place. And the timing is a bit awkward for Anthropic: the ranking shift came shortly after Elon Musk publicly praised Anthropic as a leader in the industry, which sparked its own round of discussion. Other new entries worth noting include Mark Zuckerberg’s Muse Spark 1.1, which debuted at 11th place not long after its release.

So the leaderboard is moving fast. It’s not static. And Kimi-K3’s arrival at the top is probably the clearest sign yet that Chinese AI labs are willing to compete head-on with US counterparts — not just in raw performance, but on price.

The pricing gap is genuinely large. Moonshot set Kimi-K3 at $3 per million input tokens and $15 per million output tokens. Claude Fable 5 runs at $10 per million input and $50 per million output. That’s not a small discount — it’s a roughly 70% cut on input costs and the same on output. For companies running high-volume coding tasks, that math changes budget conversations fast.

And Moonshot isn’t stopping there. The company plans to release Kimi-K3’s full model weights, with that release set for July 27. Once those weights are out, anyone can run the top-ranked coding model for free. That’s a significant move. Open weights mean developers, startups, and research teams can host the model themselves, fine-tune it, and deploy it without paying per token. It’s a different kind of competitive pressure than price-cutting — it’s basically removing the price entirely for anyone willing to self-host.

More context: Ostium Oracle Hack Drains OLP Vault of Up to $22 Million

Alibaba’s Claude Code Ban and the Bigger Picture

The competitive tension between Chinese and US AI models isn’t just playing out in leaderboard points. Alibaba recently told its staff to stop using Claude Code, citing security concerns. That directive is worth sitting with for a second. Alibaba is one of the largest tech companies in the world, and its decision to pull Claude Code from internal use — whatever the specific security reasoning — sends a signal about how Chinese companies are approaching US AI tools right now.

It’s murky whether other Chinese firms will follow suit. No details have come out on exactly what security concerns Alibaba flagged. But the move fits a broader pattern: as Chinese models get more competitive, the case for using US alternatives inside Chinese organizations gets harder to make on both performance and geopolitical grounds.

Kimi-K3’s rise sits right in the middle of all that. A model that beats Claude Fable 5 in six of seven categories, costs a fraction of the price, and will soon be available for free via open weights — that’s a pretty compelling internal argument for any organization already skeptical of US AI tools.

The AI coding space has been moving fast for years, but the pace feels different now. Models that were dominant six months ago are getting pushed down the rankings by competitors that barely existed a year before. Moonshot’s previous model was 18th. Now Kimi-K3 is first. That’s not gradual progress — that’s a hard reset on expectations.

Read also: South Korea Moves to Put Digital Assets Inside Its National Asset Management System

Arena’s rankings continue to take in votes, so the standings can shift. Kimi-K3’s lead isn’t locked in permanently. Claude Fable 5 still owns the gaming category, and Anthropic’s nine models in the top 20 give it a depth that a single first-place finish can’t match. But right now, the number at the top of the leaderboard reads 1,679, and it belongs to a Chinese lab.

Moonshot’s model weights drop July 27.