GoldieBench Blog · 6 min read
Grok Bot Mobile App + Which AI Model Scores Matter When You Review Bot Work On Your Phone, With Real GoldieBench Numbers
The phone app runs the same Bots in the cloud. These are the live leaderboard numbers that set your bar for the pages you review on it.

The grok bot mobile app has no model picker, so the GoldieBench scores worth reading are the ones that tell you what quality of finished page to expect when you review a Bot's work on your phone, and which models to run beside it in your own stack.
The Grok Bot FAQ is direct: Grok Bot picks the model for each task, there is no picker, and the choice can differ between requests.
The phone app itself is a client for Bots that work on a shared cloud computer, per xAI's mobile docs.
Your handset's chip does not change the output.
Every score below was read on live goldiebench.com pages on 13 November 2026, with the page named beside it.
If you'd like my team to work out where agents fit your SEO, book a free SEO strategy session.
Why the phone is not what you benchmark
The app handles chat, voice, approvals, computer take-overs and routine controls.
The heavy work runs in the cloud, and xAI's docs say it keeps going when the app is closed.
An iPhone on iOS 18 and an Android 9 handset get the same Bot behind them.
What the phone changes is how quickly you can approve or correct a step, not how good the step is.
What a GoldieBench score actually measures
The methodology page describes one fixed prompt per task, given to every model, one shot, no retries.
The output has to be a single self-contained HTML file that runs in a browser.
That is close to the thing you check in a Bot's Library on the phone: a finished page, made in one go.
It does not measure screen control, phone apps or approval behaviour.
The Grok deep dive says so plainly too: the bench does not cover multi-turn agent work, long-context recall or API latency.
The top of the leaderboard right now
| Rank | Model | Avg /10 | Read on |
|---|---|---|---|
| #1 | Fusion (OpenRouter) | 8.59 | goldiebench.com home leaderboard |
| #2 | Claude Opus 5 | 8.27 | goldiebench.com home leaderboard |
| #3 | Hermes MoA | 8.17 | goldiebench.com home leaderboard |
| #4 | GPT-5.6 Sol | 8.16 | goldiebench.com home leaderboard |
| #7 | Grok (xAI) | 8.09 | goldiebench.com home leaderboard |
Those averages come from the home leaderboard, which ranks by average 0 to 10 score across the tasks each model ran.
The gap between #1 and #7 is half a point.
For page-style output you review on a phone, every model in that table is in the same band.
What the Grok deep dive shows for page work
The Grok benchmarks page breaks scores down by category.
In the Pages category (3 tasks), it lists Grok at 8.3 against a field average of 7.7.
Its Landing task scored 9.0, a gold medal on that brief, and its Webos task also scored 9.0.
Its lowest builds listed there are Matrix and Aurora, both at 7.0.
That same page lists a board rank of #8 of 26 ranked frontier models, while the home leaderboard showed #7 on the same day, so check the live page before you quote a rank.
Which models to pair with your phone workflow
You can't choose what Grok Bot runs, but you do choose what runs in your own Agent OS on the desk side.
Use that side for the briefs and the checking, and let the phone handle sign-off.
- For planning briefs: a top-band model from the table above, such as Claude Opus 5 at 8.27 on the home leaderboard.
- For a second opinion on a finished page: a different vendor from the one that drafted your brief, so you aren't checking a model against itself.
- For multi-model checks: Hermes MoA, which the models page describes as a panel of frontier models merged by a chair, at 8.17.
None of these scores tell you what Grok Bot routes to, and nothing here claims they do.
A phone review loop built around the scores
- At the desk, have your Agent OS write the brief and the pass/fail checklist.
- Hand the brief to a Bot and close the laptop.
- When the Bot finishes, open its Library on the phone and check the page against the checklist.
- Reply with fixes by voice or text, or approve the next step from the card.
- Back at the desk, run the page past your second-opinion model.
Approval cards and Custom Rules are both available on the phone, per the approvals docs.
How to read a score before trusting a page
A score is an average of one-shot builds, not a promise about any single page.
Grok's own deep dive shows a spread from 9.5 on Twilightvale down to 7.0 on its weakest builds, on the same rubric.
That spread is the useful part.
It tells you even a top-band model ships an occasional weaker page, so the review step on your phone still matters.
Check structure first, then content, then the small details, and send fixes back in the same chat.
Head-to-head records on the deep dive help too: it lists Grok trailing Claude Opus 5 by 15 to 28 and leading GLM-5.2 by 24 to 9.
Mobile facts that affect how you review
- iPhone needs iOS 18 or later and Android needs 9 or later; the iOS app runs on iPad with iPadOS 18 or later.
- You can open the shared computer from a chat to watch work and take over for a password or code, then tap I'm done.
- Renaming a page needs the desktop app; to change a page's content, ask the Bot.
- Testing a routine needs the desktop app too.
- With the desktop app in use on macOS or Windows, alerts go there first, as the notifications page explains.
What the bench can't tell you about the app
It can't rate the share sheet, Siri shortcuts or the Action Button.
It can't rate approvals or phone-app tools like Calendar and Alarms.
It can't rate how fast the cloud computer finishes a long task.
Use the bench to set your quality bar for finished pages, and the docs to understand the app.
For the Android side specifically, our Grok Bot Android model guide goes deeper.
The short version
The phone doesn't change the model and you can't pick one.
The bench shows the top seven models within about half a point on one-shot builds, with Grok at 8.3 in Pages on its own deep dive.
Set your review checklist from that, and run your second opinion from the desk.
For the full Agent OS and weekly calls on model pairing, join the AI Profit Boardroom.
Also On Our Network
- Agent Operatorsa nine-step phone setup, from store listing to your first take-over
- AI Income Deskthree owner use cases, plan names and the risks before going mobile
- AI Tool Verdicta nine-criteria scorecard of the phone app, rated from the docs
- agentos.guidefour copy-paste prompts for running Bots from your phone in an Agent OS
Set your quality bar from the bench, review from your phone, and if you want my team to plan where agents fit your SEO, book a free strategy session and pair it with the grok bot mobile app.
FAQ
Can I choose the AI model in the Grok Bot mobile app?
No. The Grok Bot FAQ says Grok Bot picks the model for each task, there is no model picker, and the choice can differ from one request to the next.
What did Grok score on GoldieBench?
The home leaderboard showed an 8.09 average at #7 on 13 November 2026. The Grok deep dive lists 8.3 in the Pages category against a field average of 7.7.
Does a faster phone give better Bot results?
No. The work runs on a shared cloud computer, so the phone only changes how quickly you can answer or approve.
Does GoldieBench test the mobile app itself?
No. It scores one-shot, single-file builds from a fixed prompt. It does not test approvals, phone apps or multi-turn agent work.


