Kimi K3 Just Took the Tentative #1 Spot

Okay, this officially moved the board 🌙👑 Kimi K3 has climbed to the tentative #1 spot in our Vibe Coding Index with a 67.3, narrowly passing Claude Fable 5 at 67.0 and GPT-5.6 Sol at 66.5.

This is not a random hype bump and it is not us handing out a participation trophy. New live evidence landed, the calculation refreshed, and Kimi did enough to take the crown on the best data we have right now.

What changed

The live Arena Code WebDev leaderboard now has Kimi K3 at #1 with a 1,679 ±17 score from 1,757 votes. Claude Fable 5 is at 1,631 ±13 and GPT-5.6 Sol is at 1,618 ±13.

That matters because Arena is not another vendor slide. People compare model outputs on real front-end development prompts and vote for the result they prefer. It is a direct signal for the kind of visual, iterative product work vibe coders actually care about.

Our benchmark refresh keeps the raw Arena score, vote count, uncertainty, and status. It converts the result to a fixed 0–100 signal and blends it with the broader coding evidence already on each model profile. Price and popularity never inflate quality.

Arena still labels Kimi Preliminary, so our placement is tentative too. That label is staying right beside the score.

How Kimi passed Fable

Before this Arena result, Kimi's launch-period Coding score was 76.9 and Fable's broader Coding score was 76.5. That was only a 0.4-point edge.

The new Arena signal is much more decisive. Kimi's 1,679 becomes a normalized 97.0 in our fixed scale. Fable's 1,631 becomes 90.1. After that evidence is blended with the existing coding data, Kimi moves to 83.9 in Coding while Fable lands at 81.3.

That 2.6-point coding advantage is the entire story. Kimi still trails Fable in Intelligence, 58.8 to 59.9, and Agentic capability, 50.7 to 52.8. Fable remains the more balanced profile. However, coding carries enough weight in a practical coding index for Kimi's stronger implementation result to pull its final VCI to 67.3 versus 67.0.

Basically, Fable did not suddenly get worse. Kimi produced a stronger new signal exactly where this leaderboard is designed to care most.

There is another detail I do not want to skip. On Arena's displayed uncertainty bands, Kimi's lower edge is 1,662 while Fable's upper edge is 1,644. The current separation is not just a one-point rounding accident. More votes can absolutely move it, but the live result gives Kimi a real lead today.

Kimi versus GPT-5.6 Sol

Sol is still right there, but Kimi currently has the stronger top-line profile for this index.

• Intelligence Kimi 58.8 versus Sol 57.7
• Coding Kimi 83.9 versus Sol 81.8
• Agentic Kimi 50.7 versus Sol 51.8
• Vibe Coding Index Kimi 67.3 versus Sol 66.5

Sol keeps a small lead in long-running agentic execution. Kimi answers with a 2.1-point coding advantage and a 1.1-point intelligence advantage. That is enough for an overall 0.8-point VCI lead.

The price makes this even more interesting

Kimi is not only tentatively first. It is also the least expensive model in the top three on our current profiles.

• Kimi K3 costs $3 per million input tokens and $15 per million output tokens
• Claude Fable 5 costs $11 input and $55 output
• GPT-5.6 Sol costs $5 input and $30 output

Using our standard 3-to-1 input-to-output blend, that works out to $6 for Kimi, $22 for Fable, and $11.25 for Sol. Kimi is about 73% cheaper than Fable and 47% cheaper than Sol on that blended comparison. All three sit around a one-million-token context window.

So the current tentative winner is not asking you to pay a frontier-model tax for the crown. That changes the conversation from "Can Kimi compete?" to "How much work can Kimi replace at this price?"

What appears to be different

The safest conclusion is that Kimi is landing especially well on reference-driven web development and visual implementation. Moonshot built K3 with native visual understanding and a huge context window, which should be useful when a task combines screenshots, requirements, and a large codebase.

I am deliberately saying appears. Arena proves that voters preferred Kimi's outputs in the current WebDev pool. It does not publish a task-by-task causal breakdown proving which internal feature created every win. The visual architecture is a credible explanation to investigate, not a fact I am going to manufacture after seeing the score.

The honest status

Kimi is #1 on our current leaderboard and still tentative. Both things are true.

Arena marks the entry Preliminary. Kimi also has fewer Arena votes than Fable and Sol, and parts of its Intelligence and Agentic profile still rely on launch-period reported data while independent evidence catches up. The scheduled benchmark refresh will keep checking the source twice a day. More votes, reproduced evaluations, or new official indices can move the board again.

That is exactly how this should work. The crown moves when the evidence moves. I am not calling Kimi permanently unbeatable. I am saying that on the most current evidence available today, Kimi K3 earned the tentative #1 spot. 👑🌙

See the full leaderboard · Open the Kimi K3 scorecard · Read our methodology