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

GoldieBench deep dive

Claude Sonnet 5's full benchmark breakdown.

The agentic SWE frontier — 82% SWE-bench Verified, Dev Team mode. 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
47
Reading time
10 min
External sources
3

01 · The headline numbers

GoldieBench average
7.01/10
47 scored one-shot tasks
Board rank
#14
of 17 ranked frontier models
Task medals
1🥇 1🥈 5🥉
outright wins on shared briefs
Context
1,000,000 tokens
$3 / $15 per M ($2/$10 intro)

02 · Every benchmark, bar by bar

All 47 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

“Beautiful multi-arm spiral with convincing color gradient (warm core to violet edges), bright glowing bulge, and background starfield—clearly on-brief and polished. Full swirl/orbit/zoom interactivity with a mouse-influence vortex on the particles makes this a task winner.”

- the judge's verdict, unedited

“Renders a genuine progressively-converged Cornell box with correct red/green colored-wall bleed, a diffuse yellow sphere, and convincing glass and metal spheres showing refraction/reflection at 163 samples — physically-plausible and clearly on-brief. Minor grain and the truncated display material are the only weak points; strong direct-light NEE plus Russian-roulette path tracing make this a task winner.”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

Boards that hide the weak rows are brochures. These are Claude Sonnet 5's lowest scored builds, verdicts unedited:

“Screenshot shows only the 'Loading Solar System...' overlay on a black background — the actual scene never rendered (likely a CDN/init failure or the loading div was never hidden in the captured frame). Despite reasonable-looking source with orbits, labels, rings and controls, what renders is effectively non-functional.”

- the judge's verdict, unedited

“Only the AURORA title and hint text render over a black void — the Three.js scene (aurora, stars, mountains, moon) never appears, likely a fade overlay or render failure. The source is ambitious and well-structured but the actual result is essentially non-rendering.”

- the judge's verdict, unedited

“Screenshot captured mid-load showing only the 'ENTERING WORMHOLE...' splash on black — no tunnel visible, so the render fails to demonstrate the effect. The source is competent (TubeGeometry curve, scrolling rainbow texture, particles, mouse steering) but the 900ms loading overlay masked the actual scene at capture time, tanking the visible result.”

- the judge's verdict, unedited

“UI overlays (title, kills, weather, HP bar, hint) render correctly but the 3D scene is completely black — no terrain, trees, player, or enemies visible, meaning the WebGL world failed to render despite solid source code. A non-rendering core makes this effectively broken for a 3D RPG task.”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (23 tasks)

Claude Sonnet 57.1
field avg7.1

Others (0 tasks)

field avg7.4

Pages (3 tasks)

Claude Sonnet 57.3
field avg7.5

Sims (12 tasks)

Claude Sonnet 56.6
field avg6.7

Visuals (9 tasks)

Claude Sonnet 57.3
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 Pro63.2%www.anthropic.com
Terminal-Bench 2.180.4%www.anthropic.com
OSWorld-Verified81.2%www.anthropic.com
Humanity's Last Exam57.4% with tools - 43.2% withoutwww.anthropic.com
GDPval-AA v21618 Elo - first Sonnet to outscore the concurrent Opus flagship on any benchmarkwww.anthropic.com

From the source guides

MeasureResultSource guide
SWE-bench Verified82.1%/sonnet-5-vs-glm-5-2
First model past 80% on SWE-bench Verifiedone-shot GitHub-issue repair/sonnet-5-vs-glm-5-2

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 Claude Sonnet 5 perform on GoldieBench?

Claude Sonnet 5 averages 7.01/10 across 47 scored one-shot build tasks, ranking #14 of 17 ranked frontier models, with 1 task golds, 1 silvers and 5 bronzes. Every score is a real judged build you can open and play on this site.

02What is Claude Sonnet 5 best at?

Its strongest scored build is Fluid at 8.6/10. By category it averages Game 7.1, Page 7.3, Sim 6.6, Visual 7.3 on the bench.

03How much does Claude Sonnet 5 cost?

$3.00 input / $15.00 output per million tokens; introductory $2.00/$10.00 through 2026-08-31.

Source ledger

  1. 01Official vendor site: anthropic.comwww.anthropic.com
  2. 02SWE-bench Prowww.anthropic.com
  3. 03SWE-bench Verifiedagentos.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