Get the Agent OS + join 4,000+ founders inside the AI Profit Boardroom → Join AIPB ($59/mo)

GoldieBench deep dive

Kimi K3's full benchmark breakdown.

Moonshot's 2.8T flagship — 1M context, tuned for long-horizon agent work. Every number below is either one of my own judged one-shot builds - playable on this site - or an outside result with its source linked. Nothing display-only, nothing vibes.

Data refreshed
2026-07-17
Scored tasks
50
Reading time
10 min
External sources
9

01 · The headline numbers

GoldieBench average
7.89/10
50 scored one-shot tasks
Board rank
#9
of 17 ranked frontier models
Task medals
11🥇 8🥈 8🥉
outright wins on shared briefs
Context
1,048,576 tokens — a full codebase in working memory
$3 / M in

02 · Every benchmark, bar by bar

All 50 scored tasks, best first. Hover for the judge's comment; click through for every model's build on that task. Or overlay other models in the interactive graphs.

03 · Where it wins - with the judge's own words

“Gorgeous real geodesic raytracing with a properly lensed accretion disk wrapping over the top, convincing Doppler asymmetry, photon-ring glow, and a rich nebula starfield backdrop; polished typography and clean UI push it to the top of the field. Minor stat overlap at bottom-left ('photon sphe' clipped by hint pill) is the only blemish.”

- the judge's verdict, unedited

“Gorgeous, textbook-quality WebGL fluid sim with rich swirling dye, added particle sparkle, and polished UI (gradient title, hint pill, control buttons) — the vorticity/pressure-solve pipeline and half-float fallback handling are all correct and shippable; only minor knock is the visual is close to the well-known Pavel Dobryakov reference.”

- the judge's verdict, unedited

“Nails every synthwave trope beautifully — banded sunset, receding neon perspective grid, layered mountains, glowing gradient title, floating wireframe solids, starfield, scanlines/vignette, and generative WebAudio music. Highly polished and cohesive; a clear task winner.”

- the judge's verdict, unedited

“Gorgeous flowing volumetric aurora ribbons with convincing fbm noise, layered mountains, spruce silhouettes, moon, stars and a shooting star make a genuinely atmospheric scene; the elegant typography, palette switcher and vignette give it a shippable polish that edges past the field's best.”

- the judge's verdict, unedited

“Gorgeous on-brief cyberpunk drive: glowing multicolor building windows, cyan/magenta neon road edges, a synthwave sunset moon down the vanishing point, and clean HUD with speed/distance—strongly atmospheric and polished. Minor nitpicks are the somewhat sparse traffic and generic sign legibility, but the render clearly nails the vaporwave neon-city vibe at a task-winning level.”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

Boards that hide the weak rows are brochures. These are Kimi K3's lowest scored builds, verdicts unedited:

“The UI chrome (title, hint, control panel) renders beautifully, but the entire simulation canvas is empty black — no star, planets, asteroid belt, or trails appear, so the core N-body simulation failed to render. Strong styling can't compensate for a non-functioning main feature.”

- the judge's verdict, unedited

“The UI chrome (title, buttons, hint bar) renders beautifully with polished gradient typography, but the core galaxy particle system is completely absent — the canvas is empty black and '—PARTICLES' shows the count never populated, indicating the Three.js render failed or hung. A stylish shell around a non-functional centerpiece.”

- the judge's verdict, unedited

“HUD, title screen, and crosshair are polished and on-brief, but the 3D scene failed to render — the background is pure black despite code defining a sky dome, ground, clouds, and planes, so no actual dogfight/flight is visible.”

- the judge's verdict, unedited

“Strong scaffolding — nice sky gradient, water, clouds, sand beaches, instanced trees, and clean UI — but the terrain interior renders almost entirely black, indicating a broken lighting/normals or vertex-color issue that dominates the screenshot. The visible flaw undercuts an otherwise polished procedural island explorer.”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (23 tasks)

Kimi K38.1
field avg7.1

Others (3 tasks)

Kimi K38.3
field avg7.4

Pages (3 tasks)

Kimi K37.9
field avg7.5

Sims (12 tasks)

Kimi K37.2
field avg6.7

Visuals (9 tasks)

Kimi K38.1
field avg6.7

06 · Outside signals - every row sourced

My bench measures one thing: judged one-shot builds. These outside rows measure other things - kept separate, never blended into the GoldieBench average, every value linked to where it comes from.

Vendor-published numbers

BenchmarkResultSource
Parameters2.8T MoE - 16 of 896 experts active (corrected from the launch-day 2.5T estimate)www.kimi.com
Context window1,048,576 tokens, flat pricing across itwww.kimi.com
API pricing$3.00/M input (cache-miss) - $0.30/M cached - $15.00/M outputplatform.kimi.ai
Terminal-Bench 2.188.3 (vendor launch table)www.kimi.com
FrontierSWE81.2 (vendor launch table)www.kimi.com
BrowseComp91.2% with 300K-token compaction - 90.4% raw 1M windowwww.kimi.com

Independent evaluations

BenchmarkResultSource
LMArena Frontend Code Arena#1 debut - 1679 Elo across 1,757 votes - a 17-place jump over K2.6x.com
LMArena main text board1486 Elo across 3,026 votes (coding 1530, hard prompts 1506)x.com
Artificial Analysis indicesIntelligence 57.11 - Coding 76.24 - Agentic 50.07artificialanalysis.ai
Artificial Analysis runtime62 output tok/s - 1.99s to first token - ~$0.94 per taskx.com
Vals Index v1.274.7 - Terminal-Bench 2.1 hosted view 80.899www.vals.ai

From the source guides

MeasureResultSource guide
Launch-day community benchmarks~Fable/Sol tier; Terminal Bench the standout/kimi-k3-machine

The posts themselves

07 · Head-to-head records

08 · Methodology + honest limits

Every GoldieBench score is a one-shot, single-file build from an identical prompt - no retries, no hand-fixing - rendered for real, screenshotted, and scored 0-10 by one judge model on one rubric across the whole field. Failures score as failures. What this bench does NOT measure: multi-turn agent work, long-context recall, or API latency - that is what the sourced outside rows are for. Full method: /methodology.

Frequently asked questions

01How does Kimi K3 perform on GoldieBench?

Kimi K3 averages 7.89/10 across 50 scored one-shot build tasks, ranking #9 of 17 ranked frontier models, with 11 task golds, 8 silvers and 8 bronzes. Every score is a real judged build you can open and play on this site.

02What is Kimi K3 best at?

Its strongest scored build is Blackhole at 9.0/10. By category it averages Game 8.1, Other 8.3, Page 7.9, Sim 7.2, Visual 8.1 on the bench.

03How much does Kimi K3 cost?

Launched July 16, 2026. 2.8T-param MoE (Moonshot's quickstart corrected the circulating 2.5T estimate). $3/M input on OpenRouter at launch; included at no extra cost in the Kimi coding plan (`k3` on the coding endpoint).

Source ledger

  1. 01Official vendor site: kimi.comwww.kimi.com
  2. 02Parameterswww.kimi.com
  3. 03API pricingplatform.kimi.ai
  4. 04LMArena Frontend Code Arenax.com
  5. 05Artificial Analysis indicesartificialanalysis.ai
  6. 06Artificial Analysis runtimex.com
  7. 07Vals Index v1.2www.vals.ai
  8. 08Launch-day community benchmarksagentos.guide
  9. 09The launch-day benchmark drop that framed K3 as ~Fable/Sol tierx.com

Every model's breakdown

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 4,000+ founders shipping with it every day all live inside the AI Profit Boardroom.

4,000+founders
258documented wins
38countries
$59/momonthly