Kimi K3 Benchmarks: How It Really Ranks in 2026

If you want to know how good Kimi K3 actually is — not how good Moonshot says it is — you have to read the benchmarks carefully. This guide collects every headline number for Kimi K3 in one place, separates the independent scores from the vendor-reported ones, and flags the one caveat that most launch coverage glossed over. The short version: Kimi K3 is a top-tier model for code and reasoning, and still a step behind the leaders on agentic software engineering.

A note for builders: benchmarks describe average behavior on fixed tests rather than your particular workload. If you would rather check the numbers for yourself, OrcaRouter can bring the models together behind a single endpoint — which makes it easy enough to run Kimi K3 beside Claude, GPT or Gemini before you commit.

The independent scores (read these first)

Independent evaluators matter more than a launch slide because the vendor does not choose the test. Two independent boards covered Kimi K3 on day one:

Benchmark	Kimi K3	Context / rank	Source
Artificial Analysis Intelligence Index	57.1	#4 config, ~#3 family	Artificial Analysis
AA Coding Index	76.2	—	Artificial Analysis
AA Agentic Index	50.1	—	Artificial Analysis
Arena frontend-code rating	1679	#1, above Fable 5 (1631)	Arena / LMArena
GDPval-AA	~1668 Elo	up from K2.6's ~1190	Artificial Analysis
AA-Briefcase	1527	#2	Artificial Analysis

On Artificial Analysis, the 57.1 Intelligence Index put Kimi K3 behind Claude Fable 5 (with Opus 4.8 fallback, ~59.9) and GPT-5.6 Sol Max (~58.9), but ahead of Claude Opus 4.8, GPT-5.5 xhigh, Sonnet 5 and GLM-5.2 — a strong debut for any model, and a remarkable one for an open-weight release (source: Artificial Analysis).

The frontend-code story

The most eye-catching result is frontend code. Kimi K3 debuted at #1 on the Arena frontend-code leaderboard with a 1679 rating, above Claude Fable 5's 1631, and won six of seven categories — Fable 5 held on only in Gaming. That is a leap from its predecessor K2.6, which sat at #18 (source: Arena / LMArena). Community testers reinforced it: one-shot three.js scenes, a single-file HTML Minecraft clone, and voxel builds that testers said matched or beat Claude's output.

Vendor-reported benchmarks (and the caveat)

Moonshot's own numbers are strong — but self-reported benchmarks let the vendor pick the tests and the harness, so read them with that in mind:

Benchmark	Kimi K3 (vendor)	Comparison point
GPQA Diamond	93.5%	strongest open model reported
BrowseComp	91.2%	vendor-reported
HLE (with tools)	56.0%	vendor-reported
Terminal-Bench 2.1	88.3	vs Fable 5's 84.6

The caveat: on agentic software engineering, the picture flips. Independent and aggregate figures put Kimi K3 at FrontierSWE 81.2 (vs Claude Fable 5's 86.6) and DeepSWE 67.5 (vs 70.0) — behind the leaders. And on Terminal-Bench, tester @ChrissGPT warned that Moonshot's marketing sometimes cites Terminal-Bench 2, not the harder 2.1, so a headline "Terminal-Bench" number may not be comparing like with like. When you see a Terminal-Bench figure, check which version it is.

How to read these numbers

1. Trust independent boards over launch slides. Artificial Analysis and Arena did not let Moonshot pick the tests; the vendor table did.

2. Match the benchmark to your job. Kimi K3 leads on frontend code and reasoning (GPQA), lags on agentic SWE. Weight the one you actually ship.

3. Watch the harness. "Terminal-Bench 88.3" means little unless it is 2.1 and run the same way as the model you are comparing it to.

4. Factor in verbosity. Kimi K3 outputs roughly 2x the peer-median token count on comparable tasks, so its real cost-per-answer is higher than the per-token price suggests.

The takeaway

On the benchmarks that a third party controls, Kimi K3 is a genuine top-three model — #1 on frontend code, top-tier on reasoning, and a strong overall Intelligence Index — while sitting a clear step behind Claude Fable 5 and GPT-5.6 Sol on heavy agentic software engineering. The honest summary: frontier-adjacent, code-strong, verbose, and priced to undercut. To see how the numbers translate to your prompts, run Kimi K3 head-to-head with the models you already use.

Disclosure: spec and benchmark figures above are vendor-reported or from Artificial Analysis / Arena and are not independently audited by us. Competitor figures are attributed to their named sources and may differ from those vendors' own numbers. Community results are individual testers' first impressions, many run against a pre-release "Kivine" checkpoint, not controlled benchmarks.