Kimi K3 tops frontend coding but flunks advanced math

Moonshot’s Kimi K3 just pulled off something no Chinese model has managed before: it grabbed the number one spot on the Code Arena: Frontend benchmark. With a score of 1,679, it edged past Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618. The benchmark ranks models based on human preference ratings, so this isn’t about raw compute or sterile accuracy metrics — it’s about what actual developers like. That’s a meaningful win, and it’s got the Western AI community paying attention.

But before anyone declares a new world order, the math scores tell a much rougher story. On FrontierMath Tier 4, the hardest expert-level problems in Epoch AI’s benchmark, Kimi K3 scraped together roughly 39 percent accuracy. That’s not just below the frontier — it’s a chasm. OpenAI and Anthropic models are reportedly hitting close to 90 percent on those same problems. The gap is so wide it suggests fundamentally different model capabilities, not just a little fine-tuning lag.

The split is fascinating because it mirrors a broader tension in AI development right now. Frontend code generation relies heavily on pattern matching, understanding design conventions, and producing visually pleasing output — areas where training data quality and human feedback loops can take you far. Advanced mathematics demands something closer to genuine reasoning, the ability to chain abstract concepts and manipulate symbolic logic across long horizons. One of these things is not like the other.

So what does this actually mean for the competitive landscape? Moonshot has clearly built something impressive for a specific, commercially valuable use case. Plenty of companies would pay for a model that consistently outputs better React components than Claude or GPT. But if you’re looking for a general-purpose reasoning engine that can tackle graduate-level proofs, Kimi K3 isn’t in the same conversation yet. The real test will be whether Moonshot can close that gap in the next iteration — or if this asymmetry is baked into their approach.