A blind test win, not a self-reported score
First, let's be clear about what this leaderboard is.
Arena's Frontend Code Arena uses the Elo rating system — same logic as chess rankings. Real developers submit real coding tasks, two anonymous models each deliver an answer, and the developer blindly picks which is better. When you vote, you don't know who you're voting for.
The leaderboard updated at dawn on July 17: Kimi K3 took 1679 points, first place. Claude Fable 5, 1631, second. GPT-5.6 Sol, 1618, third. Zhipu's GLM-5.2 Max, 1587, fourth.
The previous generation, K2.6, sat at 18th with 1515 points. One product cycle, 17 spots climbed.
Win rate tells the story better than the score
Arena also disclosed head-to-head average win rates: K3 at 76%, Fable 5 at 63%, Sol at 58%.
Across seven frontend subcategories, K3 took first in six — brand and marketing, reference replication, data and analytics, consumer products, simulation apps, content creation tools. The only loss was game development, where Fable 5 still leads.
In plain terms: on "make a usable web product," K3's human-preference score currently edges out the two most expensive closed flagships.
One caveat that has to be said: K3's score is currently tagged Preliminary, with 1757 votes accumulated — fewer than Fable 5's 2505 and Sol's 2542. Small sample, ranking not fully locked. As votes pile up, the score could shift.
Overall index is still third
Don't just look at the frontend slice.
On Artificial Analysis's comprehensive intelligence index, K3 scored 57 for third place — behind Fable 5 and GPT-5.6 Sol, but ahead of Opus 4.8 and GPT-5.5. A few hard coding benchmarks:
- Terminal Bench 2.1: 88.3 (Fable 5 at 84.6)
- Program Bench: 77.8 (Fable 5 at 76.8)
- SWE Marathon: 42 (Fable 5 at 35)
In long-horizon coding, K3 does beat Fable 5. But on overall experience, Moonshot's own blog admits it: still trails Fable 5 and Sol.
Hands-on: stunning 3D, speed is the drag
Within 24 hours of launch, hands-on tests flooded X. Three high-frequency test directions: 3D interactive generation, frontend coding, visual design. Developer verdicts are remarkably consistent —
- 3D and visual generation stand out most
- Frontend UI design approaches Fable 5
- Coding is strong but hasn't surpassed the strongest closed models
- Speed is the biggest weakness
For the same task, K3's generation time runs roughly two to three times Sol's. A waterfall scene took nearly half an hour; a video edit took almost two hours.
Three rounds head-to-head: Sol wins reliability, K3 wins aesthetics
Jiemian News ran three same-prompt comparisons between K3 and GPT-5.6 Sol. The conclusion is blunt:
Round one, 3D maze game: Sol is fast and stable, a competent product that won't screw up. K3 is slower, but the visuals carry a pixel-art style, torches and wall textures have more design intent, and it's more fun to play.
Round two, cup water-pouring physics sim: K3 face-planted. First attempt, water particles phased straight through the cup wall, and when pouring, water flowed out the bottom — the physics sim didn't hold up. After a nudge it corrected, but slowly. Sol took this round.
Round three, Niagara Falls panorama: Sol was functionally complete but "phoned it in" — the rainbow the prompt asked for never appeared, and the water rendered as a slab of white particles slapped on a cliff. K3 tried harder, didn't skip the rainbow, and the water mist felt closer to real. Still slow, of course.
One-line summary: want fast and stable, pick Sol; want visual distinctiveness, pick K3 — but you'll pay two to three times the time.
Three shortcomings the company admits
Moonshot didn't hide it in their blog post:
- Sensitive to historical thinking content — switching models mid-task can visibly degrade output quality
- Training leans toward long-horizon, hard tasks — on simple everyday questions it tends to "make decisions for the user"
- Compared to Fable 5 and Sol, there's still a gap in user experience
A surprising research demo
The official showcase included a case: K3 reproducing the I-Love-Q universal relation in computational astrophysics. This typically takes a senior researcher one to two weeks. K3 did it in about two hours — read and cross-checked 20-plus papers, built a complete numerical pipeline, evaluated over 300 equations of state, found inconsistencies in published formulas, and produced 3000-plus lines of Python plus an interactive HTML dashboard.
Another dev built a CS:GO × Portal clone with K3 in three rounds of dialogue, burning about 600K tokens — at API pricing that came to just $3.24.
How to read it
The impression K3 leaves: the capability ceiling is high, but the experience hasn't fully caught up.
Topping the frontend blind test is real — blind voting cuts brand bias, developers genuinely voted with their feet. But the speed, the physics-sim first-pass accuracy, and that Preliminary tag all remind you not to crown it too fast.
July 27, when the weights drop, is the real exam. Right now, some of these scores carry a "new-model novelty premium."