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

GoldieBench deep dive

LongCat-2.0's full benchmark breakdown.

The open 1.6T MoE that builds — a frontier coder trained on non-Nvidia ASIC superpods. 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
4
Reading time
4 min
External sources
4

01 · The headline numbers

GoldieBench average
8.12/10
4 scored one-shot tasks
Board rank
#4
of 17 ranked frontier models
Task medals
0🥇 0🥈 1🥉
outright wins on shared briefs
Context
1,000,000 tokens (LongCat Sparse Attention)
Open weights · free web chat · API

Provisional flag. Part of this model's run hit infrastructure failures that are being re-run; its average is still settling. Scores shown are real judged builds only.

02 · Every benchmark, bar by bar

All 4 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

“One-shot 15KB three.js snow open-world — snow-capped mountains + 30 low-poly pines, 3000-particle falling snow, first-person glowing sword, fog. Real WASD+mouse+sprint controls, terrain-follow. verify-move: walks+looks, canvas 1440x810, 0 errors. Flawless first try — no patch.”

- the judge's verdict, unedited

“One-shot 23KB open-world explorer (the richest of the four) — rolling displaced terrain, snow mountains, a stone watchtower, 20+ conifers, boulders, grass, clouds, and terrain-height following. Real WASD+mouse. verify-move: walks+looks, 0 errors.”

- the judge's verdict, unedited

“One-shot 9KB torch-lit stone dungeon corridor — pillars, barrels, a chest, 6+ flickering torch PointLights, fog. Real WASD+mouse controls. verify-move: walks+looks, 0 errors. Lit + atmospheric (a touch over-bright orange).”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

Boards that hide the weak rows are brochures. These are LongCat-2.0's lowest scored builds, verdicts unedited:

“One-shot 9KB Minecraft-style voxel world — 16x16 grass/dirt/stone cubes, voxel trees, day/night sky, raycast break+place, real WASD+mouse. verify-move: walks+looks, 0 errors. Built the full world one-shot but the initial camera yaw faced away (sky-only) — a one-line framing patch (yaw 0, pitch -0.5) pointed it at the terrain.”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (4 tasks)

LongCat-2.08.1
field avg7.1

Others (0 tasks)

field avg7.4

Pages (0 tasks)

field avg7.5

Sims (0 tasks)

field avg6.7

Visuals (0 tasks)

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
SWE-bench Pro59.5 - edges GPT-5.5's 58.6www.marktechpost.com
Terminal-Bench 2.170.8www.marktechpost.com
SWE-bench Multilingual77.3www.marktechpost.com
BrowseComp79.9www.longcatai.org
Verification statusall vendor-reported from Meituan's internal suite; independent confirmation pendingwww.longcatai.org

From the source guides

MeasureResultSource guide
Terminal-Bench 2.170.8/longcat-2-0
SWE-bench Multilingual77.3/longcat-2-0
BrowseComp79.9/longcat-2-0
GPQA-diamond88.9/longcat-2-0
IFEval90.0/longcat-2-0

07 · Head-to-head records

Not enough shared scored tasks yet.

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 LongCat-2.0 perform on GoldieBench?

LongCat-2.0 averages 8.12/10 across 4 scored one-shot build tasks, ranking #4 of 17 ranked frontier models, with 0 task golds, 0 silvers and 1 bronzes. Every score is a real judged build you can open and play on this site.

02What is LongCat-2.0 best at?

Its strongest scored build is Dragonrealm at 8.5/10. By category it averages Game 8.1 on the bench.

03How much does LongCat-2.0 cost?

LongCat-2.0 is open-sourced (weights on Hugging Face + GitHub) and served via the longcat.chat web chat plus an OpenAI-compatible API (model id 'LongCat-2.0' at api.longcat.chat/openai/v1). It's a 1.6T-parameter MoE with ~48B activated per token, trained entirely on AI ASIC superpods (>50K accelerators, 35T+ tokens, no rollbacks). Note: the direct API key we were handed shipped with zero token quota ('Token 额度不足'), so every build here was run thr

Source ledger

  1. 01Official vendor site: longcat.chatlongcat.chat
  2. 02SWE-bench Prowww.marktechpost.com
  3. 03BrowseCompwww.longcatai.org
  4. 04Terminal-Bench 2.1agentos.guide

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