Grok Bot Windows + Which AI Models To Pair With It, With Real GoldieBench Scores

Grok Bot Windows doesn't let you pick a model, so the GoldieBench scores that matter are for the models you run around it on your PC: one to plan the brief, one to check the output, and an optional local model for private files.

The Grok Bot FAQ says Grok Bot chooses the backend model per task, with no model picker, and that the choice can change between requests.

I read every score below on the live goldiebench.com pages on 2 November 2026, and each one links to its page.

For help building a model stack around your SEO work, book a free SEO strategy session with my team.

A Windows PC runs the app, never the model

The deployment page says Bots work on a hosted Linux computer in the cloud and the desktop app is a client for chat, review and approvals.

So an x64 desktop and an Arm64 laptop, the two Windows builds the Get Started guide lists, get identical Bot output.

Your processor only decides which Setup .exe you install.

It starts to matter again only when you add a local model on the same PC.

How to read a GoldieBench score for this job

Per the methodology page, each task is one fixed prompt, one shot, with a single self-contained HTML file as the result.

The leaderboard ranks models by Julian's 0 to 10 scores averaged across tasks, and medals go to the top three per task.

That's a test of planning and shipping a working build in one pass.

It isn't a test of browser control or of whichever model Grok Bot routes your task to, so treat it as a guide for your planner and checker picks.

The leaderboard models worth pairing

Ranks, averages and medals come from the home leaderboard.

Fusion holds #1 at 8.59/10 there, but its card describes a multi-model panel, Fable 5 plus GPT-5.5 ensembled, so it's a blend rather than one model you'd brief.

I keep the pairing picks to single models so you know exactly what wrote your brief and what checked the result.

Planner pick for Windows users: Claude Opus 5

Opus 5 has the most gold medals in that table, 13, and the second-highest average on the board.

Its model page lists a 1,000,000-token context window, enough to hold a long Bot brief plus the reports it produced last week.

Use it to write the Bot's first job message and the rules for where it should stop and ask you.

Then dictate the finished brief into Grok Bot with Ctrl+D, the Windows shortcut from the changelog.

Checker pick: GPT-5.6 Sol for prompt-injection defence

A Bot that browses the web reads pages it didn't write, and xAI's approvals docs say text from a page or email is information, not your request.

A second model reviewing that output should resist planted instructions too.

The GPT-5.6 Sol page lists a prompt-injection defence score of 0.910, up from 0.697, as an outside reference benchmark rather than a GoldieBench medal score.

It also averages 8.16/10 on the bench with a 1,050,000-token context, so it can read a whole week of Bot output in one go.

What Grok's own scores say, and don't say

The Grok page shows 8.09/10 across 47 tasks with a 256,000-token context.

The Grok 4.7 page shows 7.15/10 across 20 tasks with a 500,000-token context, and lists xAI-sourced outside results including Terminal-Bench 4.0 at 38.0% and DeepSWE v1.1 (high effort) at 71.0%.

The Grok 4.7 deep dive, refreshed 27 October, puts it at board rank #18 of 26 ranked frontier models, one place lower than the #17 the home leaderboard showed when I read it.

None of that tells you which model Grok Bot uses for your task, because the FAQ says the app doesn't show it.

Why the terminal number matters on a PC

Terminal-Bench is the score closest to running shell commands, and Grok 4.7's page calls it the weak spot xAI reports itself.

On Windows, the setting that governs shell access is Execution on this computer under Settings → Computer → Your computer.

It defaults to Ask every time, and xAI recommends Never allow unless a Bot needs your local files.

Leave it there and let a stronger planner write any PowerShell you actually need, then run it yourself.

The FAQ's routing example: Opus 5.5

The Grok Bot FAQ uses Claude Opus 5.5 as its example of a model a task can be routed to.

On its GoldieBench page, Opus 5.5 averages 7.57/10 across 50 tasks with a 1,000,000-token context.

The page also says all 50 one-shot builds rendered with zero console errors and all 23 games passed the input playtest.

That's a useful sanity check, not a promise of what your Bot runs.

Your usage views won't confirm it either: the FAQ says activity is recorded as Grok Bot usage, not by provider or model name.

Local models on a Windows PC

The local models page ranks strictly on-device coders, with nothing leaving the machine.

The top local pick runs on Apple MLX, so on a Windows PC look at the Ollama entries instead.

Agents-A1 leads those at 4.83/10, well behind the cloud planners above, so keep local models for private drafts, not for writing Bot briefs.

A three-model stack around Grok Bot on Windows

1. Plan with Claude Opus 5 and paste or dictate the brief into the Windows app.
2. Run the job in Grok Bot, which picks its own model on the cloud computer.
3. Check the output with GPT-5.6 Sol before you approve anything it wants to send.
4. Keep private drafts on the PC with an Ollama-run local model.

Since v0.71.2, Bot replies land as notifications on your Windows PC while you're using it, which keeps this loop at your desk.

Using a Mac as well? Our Grok Bot Mac model guide covers the MLX side, and the AI Profit Boardroom trainings show the full stack in action.

Also On Our Network

• Agent Operatorsa seven-step install walkthrough for x64 and Arm64 PCs, silent install included
• AI Income Deskwhether the Windows app pays off for a shop, a service firm or an agency
• AI Tool Verdictan 8-criterion scorecard for the PC app, from Arm64 builds to update control
• agentos.guidefive copy-paste prompts that pair the Windows app with your local agent
• aiprofitboardroom.comthe complete Grok Bot Windows guide on the AI Profit Boardroom blog

Pick your planner and checker from the tables, run them around your next Bot job, and book a free SEO strategy session if you want my team to tune the stack for grok bot windows.