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

The Grok Bot browser doesn't let you choose the model behind it, so the real model question is what you pair around it: the planner that writes its brief, the checker that reads its output, and the agent that runs your long local jobs.

The Grok Bot FAQ says there's no model picker and Grok Bot chooses the backend model for each task.

Every GoldieBench score below was read from the live goldiebench.com pages on 2 November 2026, and each names its page.

If you want my team to pick the stack with you, book a free SEO strategy session.

What the browser does, so you know what to pair it with

The browser lives on the Bot's own cloud computer, per the computer and apps docs.

It fills logins from Vault, 1Password or a Secure Form, and the Bot never sees them.

It shows plain-word approval cards, turns pages into PDFs, keeps the original names on downloads and captures tall pages at full size.

Since v0.78.1 on the Grok Bot changelog, it enters text one letter at a time and moves the pointer on each site it fills.

None of that is model choice; it's product behaviour.

Your models sit on either side of it.

What a GoldieBench score measures here

The methodology page says every task is one fixed prompt, one shot, and the result must be a single self-contained HTML file that runs in a browser.

If it doesn't work when you drop it into a browser, it doesn't count.

So the bench is a test of building for the browser, not of driving one.

Read scores as a signal of how well a model plans and ships in one pass, which is what you want from the model writing your Bot's briefs.

Grok on the bench

The GoldieBench home leaderboard lists Grok at #7 with an 8.09 average, 43 of 47 tasks scored, and 5 gold, 1 silver and 1 bronze medals.

The Grok model page shows a 256,000-token context.

The Grok 4.7 page shows 7.15/10 across 20 tasks, with 0 gold, 2 silver and 3 bronze and a 500,000-token context.

The Grok 4.7 deep dive puts it at #18 of 26 ranked frontier models, with Neoncity as its best build at 8.7/10.

None of this tells you which model Grok Bot routes a browser task to.

Grok 4.7's agent numbers and its weak spot

The Grok 4.7 page lists outside reference benchmarks it sources to x.ai.

DeepSWE v1.1 (high effort) is 71.0%, CursorBench 4.0 is 46.3% and Terminal-Bench 4.0 is 38.0%.

It also lists Artificial Analysis Intelligence Index v4.3.2 at 46 and GDPval at 1,695 Elo.

The site labels these as reference numbers, not GoldieBench medal scores.

For long agent jobs, Terminal-Bench is the number to watch, and 38.0% is the lowest of the three coding-agent scores.

The checker: GPT-5.6 Sol, for pages the Bot didn't write

A browser reads text from strangers.

xAI's approvals docs say text from a web page doesn't count as your request.

Your checker model should hold that line 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.

It also shows 8.16/10 across 50 tasks, with 2 gold, 9 silver and 10 bronze.

On the home leaderboard Sol sits at #4 and is marked best at Pages, the bench's web-page tasks of landing pages and UIs.

That makes it the model I'd use to read a Bot's browser summary before acting on it.

The planner: Claude Opus 5

The Claude Opus 5 page shows 8.27/10 across 50 tasks with 13 gold, 7 silver and 5 bronze.

It's #2 on the home leaderboard.

Use it to write the browser brief: the pages to open, the folders to save into and the Ask first rules to set.

Claude Opus 5.5: the model the FAQ names

The Grok Bot FAQ uses Opus 5.5 as its routing example.

The Claude Opus 5.5 page shows 7.57/10 across 50 tasks with a 1,000,000-token context.

The same page says all 50 one-shot builds rendered with zero console errors.

For browser-facing output, a clean render is exactly what you want.

Long runs and second opinions: Kimi K3 and Hermes MoA

The Kimi K3 page shows 7.89/10 across 50 tasks, with 8 gold, 3 silver and 7 bronze, and a 1,048,576-token context.

Its tagline says it's tuned for long-horizon agent work, so use it to merge a month of browser PDFs and screenshots into one report.

The Hermes MoA page shows 8.17/10 across 47 tasks, #3 on the home leaderboard.

It sends one prompt to a panel and an aggregator writes one final answer, a good fit for cross-checking a client report built from browser captures.

A pairing table for the Grok Bot browser

Where each model sits in a real browser job

Take a weekly competitor check as the example.

Opus 5 writes the brief: which four pages to open, which folder to save the PDFs in, and which actions need an Ask first card.

Grok Bot runs the browser part on its cloud computer and saves the pages.

Sol reads the Bot's summary against the saved PDFs and flags anything that looks like an instruction planted on a page.

Kimi K3 merges four weeks of those folders into one trend note.

Hermes MoA gives the final note a second read before it goes to a client.

Only one of those steps happens inside Grok Bot, and that's the step you can't pick a model for.

Members of the AI Profit Boardroom get this exact stack wired into the installable Agent OS.

What the bench can't tell you about the browser

GoldieBench doesn't score form filling, sign-ins or clicking through a site.

It can't tell you how well the v0.78.1 typing works on a slow portal, or how often a handoff comes up.

For that, read the docs and watch your own runs from the computer preview.

How to read a gap between two scores

Opus 5 at 8.27 and Sol at 8.16 are a tenth apart, so pick whichever your stack already runs well.

A full point, like Opus 5 against Grok 4.7's 7.15, is worth acting on when you choose a planner.

Medals show outright task wins, not consistency, and briefs need consistency.

Browser settings that matter more than any score

The model can't fix a bad setup.

• Keep logins in Vault or 1Password, never in a prompt; the Vault docs say the Bot never sees a saved password.
• Add narrow Ask first rules for form submits, purchases and account changes.
• Accept the connector when the Bot offers one on a sign-in page, as it now does for Google Ads, Meta Ads, TikTok for Business and five other services.

For the wider cloud screen, see our Grok Bot computer use model guide.

Want the full stack set up for you? Join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe seven-step setup I run before the Bot's browser touches a login
• AI Income Deskwhere the Bot's browser saves a small business owner real hours
• AI Tool Verdictan eight-feature scorecard for the browser, with a clear verdict
• agentos.guidefive copy-paste browser prompts and a weekly web workflow for your Agent OS
• aiprofitboardroom.comthe complete Grok Bot Browser guide on the AI Profit Boardroom blog

Pick your checker model today, then book a free strategy session so my team can map the stack around your Grok Bot browser.