Grok 4.7 · xAI · released 21 Sep 2026 · benched on GoldieBench 22 Sep

Grok 4.7 In The Agent OS: 20 Games, One Shot, Real Scores

Grok 4.7 just dropped, and I put it straight to work inside the Agent OS.

I asked it to build twenty full games, one prompt each, and then I played every single one.

Some of them are the best one-shot games I have seen from any model this year.

Stick around, because one of them fell apart in a way you need to see before you trust it.

20 one-shot games, every one played before it was scoredneoncity · 8.7/10racing · 8.7/10arcade · 8.6/102 builds broke on loadGrok 4.7one prompt each20 gamesAverage 7.15 out of 10 · Grok 4.6 scored 5.95 on the same 20
Sources, the model, and every build ↓
§1What just dropped

Grok 4.7 landed on 21 September

Grok 4.7, in three facts21 Sep2026 · the release$2 / $6per million in / out500Ktoken context windowSame price as Grok 4.6. Bigger base model. Longer reinforcement learning.
sources: x.ai/news/grok-4-7 · openrouter.ai/x-ai/grok-4.7
§2The problem

The Vendor Chart Problem

What you get on launch dayLaunch posthand-picked numbersBenchmark %not a thing you can playHype threadnobody built anythingYour questioncan it ship MY app?Your toolsstill on last month's modelYour timegone reading chartsNone of it tells you if the thing actually works
THINKING IT? "Every new model says it's the best at coding."

They all do. The chart is the vendor's. The game is yours.

So I stop reading and start building. Twenty times.

§3Old way vs new way

Reading charts, or playing the games

Same new model, two ways to find outOLD WAYan hourread the launch postargue about a benchmarktry one prompt in a chat boxget a half-working demogive up and wait for reviewsNEW WAYone afternoonpick the model in the Agent OSfire twenty skill-infused game promptsplay every build on a real GPUlet a vision judge score the played framekeep the winners, learn from the wrecksOne of these gives you a scoreboard you can click
§4How it works

Grok 4.7 lives in the Grok Build tab

Where Grok 4.7 sits inside the Agent OSGrok Build tabin the Agent OSsidebarPick a model4.7 · 4.6 · 4.5one clickType what you wanta game, an app,a fixIt lands inyour Workspaceclick to playPick it once, it stays picked
§5Exactly how it works

Step by step, no mystery

One message's journeyYour promptplus the last24 turnsGrok Build CLI-m grok-4.7signed in on XSigned out?same model viaOpenRouterThe reply streamscode lands inthe workspaceSigned in or signed out, the same model answers
What actually happens on your machine1 · The pickersends model=grok-4.72 · The CLIgrok -p … -m grok-4.73 · Signed outroute falls back to OpenRouter4 · Same modelx-ai/grok-4.75 · A ```html blockis saved as a file6 · The Workspacelists it, previews itNothing invented, this is the route code
§6Live · a build in chat

Type it. Watch it stream.

real screen recording · Agent OS at localhost:3737 · recorded 22 Sep 2026

What you're watching: the real Grok Build tab. I click 4.5, 4.6, then 4.7, type one line asking for a neon snake game, hit send, and Grok 4.7 streams the whole game file live.

Pick the model. Type the game. It lands in your workspace.
§7Live · the workspace

Every build, one click to play

real screen recording · Agent OS at localhost:3737 · recorded 22 Sep 2026

What you're watching: the Workspace tab of the same Grok Build screen. The twenty Grok 4.7 game files are listed, I open three of them, and each one plays inside the Agent OS.

§8How I tested

20 games, every one played

The test, start to finish20 game briefsskill-infusedthree.js promptsOne shot eachno retries,no hand fixesReal GPU playMetal, ~70fps,keys + mouseHUD sampleddid the numbersmove?Vision judgescores theplayed frameA posed screenshot never counts. The played frame does.
same 20 briefs Grok 4.6 got in Augustsame harness, same judgeno hand patching, ever
§9The scoreboard · top half

Average 7.15 out of 10

Grok 4.7 · the best one-shot buildsneoncity8.7 / 10racing8.7 / 10arcade8.6 / 10dragonrealm8.6 / 10flightsim8.5 / 10neonracer8.4 / 10outrun8.4 / 10skyrim8.4 / 10twilightvale8.3 / 10neonblaster8.2 / 10Played on a real GPU, scored on the played frame
§10The scoreboard · bottom half

And the bottom half

Grok 4.7 · the restrpg8.2 / 10gtadrive7.8 / 10gtafoot7.2 / 10dogfight6.3 / 10doom6.3 / 10raycaster6.3 / 10parachute6.0 / 10nordiccrypt5.5 / 10crypt2.3 / 10voxelcraft2.2 / 10Red bars died on load or never moved
§114.7 vs 4.6

14 wins, 5 losses, 1 ties

Same 20 games, same harness, one month apart7.15Grok 4.7 average5.95Grok 4.6 average14–5head to head · 1 tiedThis is the number that matters
Grok 4.6 scored 2026-08-13 on the same twenty briefs · Grok 4.7 scored 2026-09-22
§12Best build · one

neoncity · 8.7 out of 10

real gameplay · one-shot build · filmed on Metal, no slow motion

What you're watching: Grok 4.7's one-shot neoncity, played live on a real GPU. 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 conf

Play neoncity yourself →the actual file Grok 4.7 wrote · opens on goldiebench.com
Played, not posed.
§13Best build · two

racing · 8.7 out of 10

real gameplay · one-shot build · filmed on Metal, no slow motion

What you're watching: Grok 4.7's one-shot racing, played live. Strong: gorgeous dusk-lit third-person racer with a detailed low-poly car, banked track, checkpoint gate, obstacles, working minimap and rich HUD; playtest shows real speed (105), score (000749) and lap progression plus

Play racing yourself →the actual file Grok 4.7 wrote · opens on goldiebench.com
§14Best build · three

arcade · 8.6 out of 10

real gameplay · one-shot build · filmed on Metal, no slow motion

What you're watching: Grok 4.7's one-shot arcade, played live. 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

Play arcade yourself →the actual file Grok 4.7 wrote · opens on goldiebench.com
§15Side by side · doom

Grok 4.7 · Grok 4.6 · Claude Fable 5

three real one-shot builds, filmed on a real GPU, labelled underneath · same prompt to each model

What you're watching: the Doom-style raycaster brief, built one-shot by Grok 4.7, Grok 4.6 and Claude Fable 5, all three played at real speed so balance problems show.

§16Side by side · arcade

The 3D breakout, three ways

three real one-shot builds, filmed on a real GPU, labelled underneath · same prompt to each model

What you're watching: the 3D breakout arcade brief from Grok 4.7, Grok 4.6 and Claude Fable 5, all three driven at the same time.

§17Where it broke

2 of 20 died on load

The wrecks, with the real console errorcrypthud is not a functionvoxelcraftmesh.computeBoundingSphere is not One bad line kills a forty-kilobyte game
THINKING IT? "A few crashes out of twenty isn't bad."

It isn't. It's also why you never ship a one-shot build without opening it.

Hand the model its own error and it usually fixes it on the first retry.

The vendor chart is not the game.
§18Every build

All 20, click any of them

Grok 4.7 one-shot build of neoncity — GoldieBench playtest frame
8.7neoncityone shot · played · scored
Grok 4.7 one-shot build of racing — GoldieBench playtest frame
8.7racingone shot · played · scored
Grok 4.7 one-shot build of arcade — GoldieBench playtest frame
8.6arcadeone shot · played · scored
Grok 4.7 one-shot build of dragonrealm — GoldieBench playtest frame
8.6dragonrealmone shot · played · scored
Grok 4.7 one-shot build of flightsim — GoldieBench playtest frame
8.5flightsimone shot · played · scored
Grok 4.7 one-shot build of neonracer — GoldieBench playtest frame
8.4neonracerone shot · played · scored
Grok 4.7 one-shot build of outrun — GoldieBench playtest frame
8.4outrunone shot · played · scored
Grok 4.7 one-shot build of skyrim — GoldieBench playtest frame
8.4skyrimone shot · played · scored
Grok 4.7 one-shot build of twilightvale — GoldieBench playtest frame
8.3twilightvaleone shot · played · scored
Grok 4.7 one-shot build of neonblaster — GoldieBench playtest frame
8.2neonblasterone shot · played · scored
Grok 4.7 one-shot build of rpg — GoldieBench playtest frame
8.2rpgone shot · played · scored
Grok 4.7 one-shot build of gtadrive — GoldieBench playtest frame
7.8gtadriveone shot · played · scored
Grok 4.7 one-shot build of gtafoot — GoldieBench playtest frame
7.2gtafootone shot · played · scored
Grok 4.7 one-shot build of dogfight — GoldieBench playtest frame
6.3dogfightone shot · played · scored
Grok 4.7 one-shot build of doom — GoldieBench playtest frame
6.3doomone shot · played · scored
Grok 4.7 one-shot build of raycaster — GoldieBench playtest frame
6.3raycasterone shot · played · scored
Grok 4.7 one-shot build of parachute — GoldieBench playtest frame
6.0parachuteone shot · played · scored
Grok 4.7 one-shot build of nordiccrypt — GoldieBench playtest frame
5.5nordiccryptone shot · played · scored
Grok 4.7 one-shot build of crypt — GoldieBench playtest frame
2.3cryptone shot · played · scored
Grok 4.7 one-shot build of voxelcraft — GoldieBench playtest frame
2.2voxelcraftone shot · played · scored
§19Official numbers

xAI's own chart, next to the independent one

xAI's launch-post numbersCursorBench 4.0 · Grok 4.746.3%CursorBench 4.0 · Grok 4.640.4%DeepSWE v1.1 · Grok 4.771.0%DeepSWE v1.1 · Grok 4.665.2%Their chart, their benchmarks
Artificial Analysis Intelligence Index v4.3.2Claude Fable 5.153GPT-653Grok 4.746Independent, and a seven-point gap
sources: x.ai/news/grok-4-7 · the-decoder.com, 21 Sep 2026 · artificialanalysis.ai
§20You might be thinking

“I'll wait for the reviews.” “Grok is just for X.”

Wrong: “I'll wait for the reviews.”
Right: You just watched twenty. The review is the scoreboard, and every build is a click away.
Wrong: “Grok is just the X chatbot.”
Right: Grok Build is a full coding agent with a CLI, and it now sits inside the Agent OS next to Claude, Codex and Hermes.
Wrong: “I'd need to be signed in to everything.”
Right: Signed out of Grok on X? The tab falls back to the same model through OpenRouter. It just works.
Don't take my word for it

Members post their wins every day — agency owners, ecom founders, course creators, solo operators across 38 countries.

Read the 158-page wins doc →
§21 Skip the setup

Get the Agent OS with Grok 4.7 already wired in

Come join us in the AI Profit Boardroom. The Agent OS is where Grok 4.7, Claude, Codex and Hermes share one memory and one workspace, so the game it builds today is the app it fixes tomorrow.

The Agent OS zip — the Grok Build tab with the 4.7 / 4.6 / 4.5 picker and the OpenRouter fallback
Every CLI you already pay for in one dashboard — Claude, Codex, Gemini, Kimi, GLM, Grok
The GoldieBench harness — bench any new model on real games the week it ships
Four coaching calls a week — bring your own setup and we fix it live
Daily tutorials and a prompt library
A member map to find people near you building the same thing
3,900+ business owners already inside
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
4 live calls a week · daily tutorials · a member map · used in 38 countries
§22Cost

Does it burn tokens?

What one game costs through OpenRouter$6per million output tokens~12Ktokens in a 40KB game~$0.08per one-shot game20 games cost less than a coffee
THINKING IT? "Doesn't running Agent OS burn a fortune in tokens?"

No — that's the biggest myth about it. Agent OS runs the everyday 90% on a free local model on your own machine, free APIs slot in for more, and for the frontier work it drives the CLIs you already pay for — Grok Build on your X Premium+ plan, the Claude CLI on your Claude subscription. It's a layer on top of what you already own, not a new meter. And inside the AI Profit Boardroom there are full token-optimisation tutorials, so you learn to cut usage to the bone.

§23Do this first

Open the tab. Pick 4.7. Build one game.

Your next five minutesOpen Grok Buildin the Agent OSsidebarClick 4.7it stayspickedType one gameone lineis enoughOpen Workspaceand playwhat it madeThen go check its score on GoldieBench
See Grok 4.7 on GoldieBench →every build, every verdict, the head-to-head with 4.6
§24 Your move

Get more leads and customers, with your business automated

Grok 4.7 builds the game. The Agent OS is what turns that into leads, content and a business that runs while you sleep. Readers bookmark this page. Operators join, install the Agent OS this week, and ship.

The full Agent OS — Grok 4.7, Claude, Codex, Hermes, one memory, one workspace
Free local models for the everyday 90%, paid CLIs only for the frontier work
The 30-day roadmap and the video walkthrough
Four coaching calls a week, daily tutorials, a prompt library
3,900+ founders across 38 countries
Join the AI Profit Boardroom →Inside the AI Profit Boardroom · skool.com/ai-profit-lab
258 documented member wins · 38 countries · 5 live calls a week · everything from this guide, pre-wired