GoldieBench deep dive
Grok 4.7's full benchmark breakdown.
Grok 4.6's price, a bigger base model, and it builds games that hold together. 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.
01 · The headline numbers
02 · Every benchmark, bar by bar
All 20 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
“Renders a gorgeous cyberpunk street with pink/cyan neon curbs, glowing buildings, a hovercar with reflective plate lights, and a full working HUD (speed, hull/boost/route, score, extract distance, minimap); playtest confirms real motion, progressing score and checkpoint/traffic state changes with no errors. Extremely polished and clearly on-brief — a task-topping contender.”
“A gorgeous 3D breakout with rich lighting, gradient sky dome, gridded arena, minimap, and a polished sci-fi HUD; playtest confirms real motion (walkPx 0.23) with hull/lives/score/bricks all changing and no console errors — clearly at or above the field's best.”
04 · Where it struggles - quoted, not hidden
Boards that hide the weak rows are brochures. These are Grok 4.7's lowest scored builds, verdicts unedited:
“The 3D world never rendered — the canvas stays black with only the HUD/hotbar visible, and a console error (computeBoundingSphere is not a function) plus zero pixel-change and frozen HUD confirm the voxel scene never appeared. Nicely designed HUD, but the actual game is broken/non-rendering.”
“The 3D scene never rendered — screen is pure black (lum mean 2.3, zero pixel change across all playtest moments) with a fatal 'hud is not a function' console error freezing the game at DEPTH 000 / SCORE 0000. Only the polished HUD chrome shows; the actual crypt crawler is non-functional.”
“Renders a functional first-person crawler with a nice torch, minimap, polished HUD and working movement (depth advanced 000→018m), but the brief-critical atmosphere is broken: no torch-lit dungeon mood — the scene is bright pale-blue with a daytime sky, an out-of-place camera clipping into a wall, and the score/seal never progressed during play, reading generic rather than a dark Nordic crypt.”
“Renders a polished HUD (altimeter, LZ bearing compass) and a decent low-poly jungle/skydiver with shadows, but the playtest shows the run instantly ending at 'MISSED — No canopy' with altitude at 0001 and the screen frozen (mid/late pixel-change near zero) — the freefall/deploy/steer loop never actually played out.”
05 · Category breakdown vs the whole field
Games (20 tasks)
Others (0 tasks)
Pages (0 tasks)
Sims (0 tasks)
Visuals (0 tasks)
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.
From the source guides
| Measure | Result | Source guide |
|---|---|---|
| Artificial Analysis Intelligence Index v4.3.2 | 46 | /artificialanalysis.ai |
| CursorBench 4.0 | 46.3% | /x.ai |
| DeepSWE v1.1 (high effort) | 71.0% | /x.ai |
| GDPval | 1,695 Elo | /x.ai |
| Terminal-Bench 4.0 | 38.0% | /x.ai |
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 Grok 4.7 perform on GoldieBench?
Grok 4.7 averages 7.15/10 across 20 scored one-shot build tasks, ranking #18 of 26 ranked frontier models, with 0 task golds, 2 silvers and 3 bronzes. Every score is a real judged build you can open and play on this site.
02What is Grok 4.7 best at?
Its strongest scored build is Neoncity at 8.7/10. By category it averages Game 7.1 on the bench.
03How much does Grok 4.7 cost?
xAI's September 2026 release, priced exactly like Grok 4.6: $2 per million input tokens and $6 per million output on the standard tier, $4 / $12 on the Fast tier (double the output speed). OpenRouter lists the same model at $1.60 / $4.80.
Source ledger
- 01Official vendor site: x.ai/news/grok-4-7x.ai
- 02Artificial Analysis Intelligence Index v4.3.2agentos.guide
- 03CursorBench 4.0agentos.guide
- 04Julian's guide: /grok-4-7-agent-osagentos.guide
Every model's breakdown
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 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.