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

GoldieBench deep dive

GPT-5.6 Sol's full benchmark breakdown.

OpenAI's flagship — the Sun of the 5.6 lineup. 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
5

01 · The headline numbers

GoldieBench average
8.16/10
50 scored one-shot tasks
Board rank
#3
of 17 ranked frontier models
Task medals
9🥇 9🥈 7🥉
outright wins on shared briefs
Context
1,050,000 tokens
$5 / $30 per M

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

“Strong classic matrix rain with authentic katakana glyphs, bright white leaders, tasteful pink accent streams, and clean neon HUD (title, status, hint, live counters) — the subtle grid and glow add polish while interactivity (bend/pulse) elevates it well above the field's best.”

- the judge's verdict, unedited

“Gorgeous, textbook synthwave scene—striped sun, layered mountains, city silhouette, palms, glowing pink-edged road with proper pseudo-3D curve and a neon car—all polished with excellent HUD and title treatment. Only minor nit is the somewhat abstract car sprite, but overall this matches the field's best.”

- the judge's verdict, unedited

“Gorgeous, fully-rendered neon breakout with rainbow brick grid, glowing paddle/ball, retro perspective grid floor, and clean HUD/controls/pause overlay — strong arcade identity backed by solid physics, DPR scaling, particles and audio. Polish and cohesion put it at the top of the field.”

- the judge's verdict, unedited

“Gorgeous layered green-to-violet curtains with soft blur, twinkling stars, silhouetted mountains and elegant typography make this genuinely cinematic and on-brief. Interactive wind/tap hints and Kp status polish it; only minor risk is the aurora ribbons overlapping the H1 slightly, but overall it matches the field's best.”

- the judge's verdict, unedited

“Beautifully rendered third-person chase-cam scene with a detailed player jet, layered clouds, moon, enemy squadron inbound, and a cohesive HUD (crosshair, armor/boost meters, working radar with blips). Cinematic art direction and complete combat framing make it a clear task winner; only minor risk is depth in actual combat feel.”

- the judge's verdict, unedited

04 · Where it struggles - quoted, not hidden

Boards that hide the weak rows are brochures. These are GPT-5.6 Sol's lowest scored builds, verdicts unedited:

“The HUD, water plane, and clouds render cleanly with polished branding, but the entire voxel island terrain is rendered as an unlit black silhouette — likely a lighting/vertexColor/instanceColor failure — which guts the core Minecraft-style landscape brief.”

- the judge's verdict, unedited

“The screenshot shows a broken top-down camera view (not the intended third-person on-foot perspective) — the world renders as a flat blurry grid with no visible 3D buildings, pedestrians, or player character, only a stray triangle marker; HUD elements (health, weapon, ammo, hint) are present but the core rendered scene fails the brief. Strong HUD/code ambition, but the actual render is essentially non-functional as a GTA on-foot experience.”

- the judge's verdict, unedited

“Real GLSL path tracer with diffuse/metal/glass materials, area light importance sampling and progressive accumulation is technically legit, and the diffuse pink sphere plus glass sphere read convincingly. But the render is visually broken: the floor is largely black with jagged noisy artifacts, the checker plane and refractive spheres show distracting aliasing/fireflies, and the framing feels off — polished chrome but flawed output keeps it below the shippable bar.”

- the judge's verdict, unedited

“Renders a clean stylized low-poly terrain with trees, biome/relief/sun controls, seed, compass and cruise mode — a solid, polished procedural explorer. But the scene reads flat and washed out with heavy fog obscuring the relief, so it lacks the dramatic elevation and depth of a top build.”

- the judge's verdict, unedited

05 · Category breakdown vs the whole field

Games (23 tasks)

GPT-5.6 Sol8.2
field avg7.1

Others (3 tasks)

GPT-5.6 Sol8.3
field avg7.4

Pages (3 tasks)

GPT-5.6 Sol8.5
field avg7.5

Sims (12 tasks)

GPT-5.6 Sol8.3
field avg6.7

Visuals (9 tasks)

GPT-5.6 Sol7.8
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 Pro64.6%openai.com
Terminal-Bench 2.188.8% (Sol) - 91.9% (Sol Ultra)openai.com
DeepSWE72.7%openai.com
BrowseComp90.4% (Sol) - 92.2% (Sol Ultra)openai.com
Agents' Last Exam53.6 - a new high, +13.1 over Claude Fable 5openai.com

Independent evaluations

BenchmarkResultSource
GA dategenerally available July 9, 2026 across ChatGPT, Codex and the APIsimonwillison.net

From the source guides

MeasureResultSource guide
HealthBench Professional (Sol)60.5 (+8.7 vs 5.5)/gpt-5-6-models
Prompt-injection defence0.910 (up from 0.697)/gpt-5-6-models
Internal CTF (cyber)96.7% — saturated/gpt-5-6-models

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 GPT-5.6 Sol perform on GoldieBench?

GPT-5.6 Sol averages 8.16/10 across 50 scored one-shot build tasks, ranking #3 of 17 ranked frontier models, with 9 task golds, 9 silvers and 7 bronzes. Every score is a real judged build you can open and play on this site.

02What is GPT-5.6 Sol best at?

Its strongest scored build is Blackhole at 8.8/10. By category it averages Game 8.2, Other 8.3, Page 8.5, Sim 8.3, Visual 7.8 on the bench.

03How much does GPT-5.6 Sol cost?

GPT-5.6 shipped as three models — Luna ($1/$6 per M), Terra ($2.50/$15) and Sol ($5/$30) — each with a same-price pro variant that ships a higher default reasoning effort. All share a ~1.05M-token context window and are rated High capability. Benched here on the flagship, Sol, at medium reasoning effort via OpenRouter.

Source ledger

  1. 01Official vendor site: openrouter.aiopenrouter.ai
  2. 02SWE-bench Proopenai.com
  3. 03GA datesimonwillison.net
  4. 04HealthBench Professional (Sol)agentos.guide
  5. 05Julian's guide: /gpt-5-6-prepagentos.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