GoldieBench Blog · 6 min read
Best Grok Bot Uses And Which AI Models To Pair With Them, Using Real GoldieBench Page Scores
Grok Bot chooses its own model, so the leaderboard matters for the pages and recaps you build from its work.

The best Grok Bot uses don't come with a model setting, so the GoldieBench scores that matter are the ones for the models you run beside your Bots: the pages, dashboards and recaps your routines and meetings feed.
Book a free SEO strategy session →
The Grok Bot FAQ says Grok Bot picks the model for each task, offers no model picker, and doesn't show which model handled a request.
So this page doesn't rank models inside Grok Bot. It matches each top use with the GoldieBench numbers for the work that sits next to it, all read on live GoldieBench pages on 13 November 2026.
Want help choosing a stack? Book a free SEO strategy session with my team.
The uses this page pairs models with
From the use cases page and the feature docs, the uses with the clearest output are:
- Routines: scheduled or event-triggered jobs, such as a weekday briefing.
- Meeting notes: a summary and action items after Google Meet, Zoom or Microsoft Teams calls, on Team, Enterprise and Ultra plans.
- Drafted replies: email and Slack drafts that wait for approval.
- Vault sign-in and computer use: exports from portals with no app.
- Team Bots and group chats: several Bots handing work along.
Each of those produces raw material. What you build from it, a client report page or an internal dashboard, is where your own model choice comes in.
How GoldieBench scores work
The methodology page sets the rules.
- Every model gets the same fixed prompt, one shot, with no iteration.
- Each result is a single self-contained HTML file that runs in a browser.
- Scores are Julian's own 0 to 10 numbers, and medals go to the top three per task. Ties at the top share gold.
No task on the bench is a meeting summary or an inbox sort.
The nearest category is Pages, with three tasks on the GoldieBench leaderboard: Landing, Webos and Aipbpromo.
The leaderboard top seven
Read from the leaderboard on 13 November 2026:
| Rank | Model | Average | Best at |
|---|---|---|---|
| #1 | Fusion | 8.59 | Games |
| #2 | Claude Opus 5 | 8.27 | Sims |
| #3 | Hermes MoA | 8.17 | Visuals |
| #4 | GPT-5.6 Sol | 8.16 | Pages |
| #5 | Claude Fable 5 | 8.10 | Games |
| #6 | Qwen 3.8 | 8.10 | Sims |
| #7 | Grok | 8.09 | Games |
The overall average blends games, sims, visuals and pages, so for report-style output the Pages rows below tell you more.
Book a free SEO strategy session →
The Pages scores that matter here
Three models stand out on the page-style tasks.
| Model | Landing | Webos | Aipbpromo | Source |
|---|---|---|---|---|
| Grok | 9.0 (gold) | 9.0 (gold) | 7.0 | Grok model page |
| GPT-5.6 Sol | 8.7 | 8.4 | 8.3 (silver) | GPT-5.6 Sol page |
| Task average | 8.11 | 7.98 | 7.48 | Task pages |
The Landing task page lists 26 models tested, 21 scored, an 8.11 average and Fugu Ultra as the winner. Grok's 9.0 shares gold under the tie rule.
The Webos task averages 7.98 with Fusion as winner, and Aipbpromo averages 7.48 with Qwen 3.8 on top.
Routines: pair them with a page model
A routine's output lands as a message in the Bot's chat, or on a Page if you have Pages. Pages are still rolling out and aren't on Enterprise teams yet, per the files and results docs.
If you want that weekly output turned into a page you host yourself, use a model with strong Landing and Webos numbers.
On these numbers, Grok's two 9.0 page scores make it the first pick for a layout-led weekly report page, with GPT-5.6 Sol, the leaderboard's "best at Pages" model, as the steadier all-rounder across all three page tasks.
Meeting notes: recaps your clients see
The Bot sends a summary and action items after each call.
For a recap page you send to a client, GPT-5.6 Sol's 8.3 silver on Aipbpromo, a motion-heavy promo page, suggests it handles animated, scene-based layouts better than Grok's 7.0 on the same task.
For a plain, structured recap, either model's Landing score clears the 8.11 task average.
If you also have Pages in Grok Bot, a longer meeting report goes onto a page with the link in the chat, so you may only need your own model for client-facing versions.
Portal exports: dashboards from the data Vault unlocks
With logins saved in Vault, a Bot signs in to a supplier or analytics portal on its cloud computer and downloads a file. The Vault page says the Bot never sees the saved password, card number or security code while doing it.
Book a free SEO strategy session →
That export is often a spreadsheet, and the natural next step is a small internal dashboard.
The closest bench task is Webos, a working desktop with windows, a dock and three apps, which tests interactive layouts more than static copy.
Grok's 9.0 on Webos is the higher of the two scores in the table above, against a 7.98 task average, so it's the one to try first for a click-through dashboard. Fusion, the Webos winner, is a multi-model panel rather than a single model.
GPT-5.6 Sol's 8.4 on the same task still sits above the average if you'd rather keep one model for every page.
Team Bots and group chats: the multi-model angle
A group chat lets two to six Bots pass work along, and you give each stage an owner.
GoldieBench has a similar idea at the model level. Hermes MoA, a panel of models merged by a chair, sits #3 overall at 8.17, and Fusion, a multi-model panel, leads at 8.59.
That doesn't change what Grok Bot runs. It's a pointer for your own Agent OS: when one step needs the strongest output, a panel model scores higher on average than any single model on the board. More on shared Bots is in the Grok Bot teams breakdown.
What about Grok 4.7?
The Grok 4.7 page shows a 7.15 average across 20 scored tasks, all of them games, and the leaderboard places it at #17.
With no Pages scores on that page, there's nothing on GoldieBench yet to support it for report or dashboard work.
The separate "Grok" entry, with 43 of 47 tasks scored, is the one with the page numbers above.
What these scores don't tell you
- They don't measure Grok Bot itself, which chooses its own model per task.
- They don't test meeting transcription, phone calls or Vault sign-in.
- They're one-shot builds, so a model you iterate with may do better on your own pages.
Read them as a guide for the work around your Bots, not as a ranking of the Bots.
They also change. The leaderboard counts 50 tasks and 33 models today, and new entries can move every average on it.
A simple pairing plan
- Set up a weekday routine in Grok Bot that collects the numbers you report on.
- Pick Grok or GPT-5.6 Sol in your Agent OS to turn that output into a page.
- Use Hermes MoA or Fusion for the one step where quality matters most.
- Re-check the leaderboard monthly, because new models join the bench.
For the video walkthroughs, join the AI Profit Boardroom.
Book a free SEO strategy session →
Also On Our Network
- Agent Operatorsthe ten Grok Bot setups I'd run first, in order, with every button
- AI Income Deskwhich Grok Bot jobs a shop, a service firm and an agency should hand over first
- AI Tool Verdicta scorecard rating ten Grok Bot features out of 10 from the docs
- agentos.guidefive copy-paste prompts that give Grok Bot its roles in your Agent OS
Pick one page model, connect it to your routine output, and talk to my team or join the AI Profit Boardroom to pair the right models with the best Grok Bot uses.
FAQ
Can I choose the model Grok Bot uses?
No. The Grok Bot FAQ says it picks the model for each task and has no model picker.
Which model scores best on GoldieBench page tasks?
Grok scores 9.0 on Landing and 9.0 on Webos, and GPT-5.6 Sol is the leaderboard model listed as best at Pages, with 8.7, 8.4 and 8.3 across the three page tasks.
Does GoldieBench test Grok Bot?
No. GoldieBench scores models on one-shot HTML builds. It does not test Grok Bot features like meetings or Vault.
Where does Grok 4.7 rank?
The leaderboard places Grok 4.7 at #17 with a 7.15 average across 20 scored tasks, all games.


