Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-27 · high
It tops Frontend Code Arena but trails Opus 5 by more than three points on SWE-bench Verified.
Why now: The comparison comes from Nathan Lambert's July 27 breakdown of K3's scores across four leaderboards.
K3 took the top spot on Frontend Code Arena, scoring 1,679 points to Fable 5's 1,631, per Nathan Lambert's analysis of the model.
It's the first open-weights model to lead that board, and it wins six of the seven frontend domains tested, losing only Gaming. For a team weighing an open model against a closed one for UI-generation work, that's a concrete number to work with.
The win doesn't carry over to general engineering. On SWE-bench Verified, which tests work across a whole codebase instead of one interface, K3 scores 93.4% against Opus 5's 97.0%. So the frontend lead doesn't generalize past that one benchmark category.
K3 also ranks #2 on the Vals AI index and #3 on Artificial Analysis, where it trails only Fable 5 and GPT-5.6 Sol.
Lambert credits Kimi Delta Attention paired with Attention Residuals, layered onto a scaled mixture-of-experts design, for roughly 2.5x better scaling efficiency than K3's predecessor, K2.
That's the real result here. An open-weights model can out-code closed models on one visual task and still lose the broader engineering fight by more than three points. The open-versus-closed debate isn't one scoreboard anymore. It's benchmark by benchmark. Check which board matches what you're building before you crown anything the new open-weights leader.
Each link below shares sources, entities, or timing with this story.
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover Attention Residuals, Fable, GPT, Kimi Delta Attention; overlapping topics (attention, code, fable, frontend, model).
Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Claude Code uses Opus); both cover MoE, Opus, Verified; overlapping topics (code, coding, model).
Simon Willison uses Fable / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Simon Willison uses Fable); both cover Artificial Analysis, Fable, GPT, Opus; overlapping topics (fable, model).
Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Claude Code uses Opus); both cover Fable, GPT, Opus; overlapping topics (code, coding, fable).
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Opus built by Anthropic); both cover GPT, MoE, Opus; overlapping topics (coding, model).
Linked by a graph relationship (Opus built by Anthropic); both cover GPT, MoE, Opus; overlapping topics (coding, model).
Opus built by Anthropic / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Opus built by Anthropic); both cover Fable, GPT, MoE; overlapping topics (board, coding, model).
Claude Code uses Opus / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Claude Code uses Opus); both cover Fable, GPT, MoE, Opus; overlapping topics (behind, code, fable, model).