Grok Bot Teams + Which AI Models Matter: What Real GoldieBench Scores Tell A Team Using Shared Bots

Grok Bot teams doesn't let you choose a model, so GoldieBench scores matter in two places: judging xAI's own models, and picking models for any open team stack you run beside it.

Team Bots shipped in Grok Bot v0.61.0 on 26 September 2026, and the docs now call them Shared Bots: one Bot, shared with the team, with a private chat for each person.

The Grok Bot FAQ is blunt about models: there's no model picker.

So this page reads the live GoldieBench board, read on 31 October 2026, and says what each number does and doesn't tell a team.

Want help choosing a stack for your team? Book a free SEO strategy session with my team.

What the Grok Bot docs say about models

Grok Bot picks the model for each task, and you can't route a request to or away from a given model.

The FAQ says the choice can be a first-party SpaceXSI model or a third-party model from a provider on the sub-processor list, and it can differ between requests.

It also says Grok Bot doesn't show which model handled a request.

For teams, one more line matters: your team's Cursor model allowlist or blocklist doesn't apply to Grok Bot.

The teams and enterprises page backs this up with "Grok Bot manages model selection."

So no benchmark score will change what a shared Bot runs on.

What a score can do is set expectations and guide the models you run elsewhere.

GoldieBench scores for xAI's own models

These are the xAI rows on the GoldieBench leaderboard, quoted as the site shows them.

The Grok model page lists a 256,000-token context and 5 gold, 1 silver and 1 bronze medal.

Its best Page tasks were Landing Page and Webos, both 9.0 with gold.

The Grok 4.7 page lists a 500,000-token context and xAI-reported numbers: GDPval 1,695 Elo, DeepSWE v1.1 71.0%, CursorBench 4.0 46.3% and Terminal-Bench 4.0 38.0%.

That same page notes 2 of 20 Grok 4.7 builds died on load or never moved when played.

What these numbers mean for a shared team Bot

GoldieBench measures one-shot builds: the same fixed prompt, one try, one self-contained HTML file, scored 0 to 10, per the methodology page.

A shared Bot's daily work is different: answering teammates, using apps, filling logins and posting in Slack threads.

So read the Grok rows as a signal of xAI's model quality, not a forecast of your shared Bot's answers.

The Grok 4.7 GDPval figure is the closest outside signal to office work on that page, and it's xAI-reported.

And because Grok Bot may route a task to a third-party model, even the xAI rows aren't guaranteed to be what runs.

The useful takeaway is simpler: the work quality depends on your instructions, files and approvals more than on a model you can't pick.

If you want to pick the model: the top of the board

Some teams want a shared assistant where they control the model.

That means an open stack beside Grok Bot, and GoldieBench is a fair place to start shortlisting.

All five rows come from the leaderboard on 31 October 2026.

Grok sits 0.08 behind Hermes MoA and 0.50 behind Fusion on the overall average.

Why Hermes MoA fits a team setup

The Hermes MoA page describes a panel of frontier models whose drafts a named aggregator merges into one answer.

Its default panel is Claude Opus 4.8 plus GPT-5.5, aggregated by Opus 4.8, and every slot can be swapped from the Mixture tab in the Agent OS.

It scores 8.17 across 47 scored tasks.

The trade-off on the same page is speed: about 110 to 140 seconds per single-file build, since the panel and the aggregator are separate calls.

For a team, that's the opposite trade to Grok Bot.

You get model choice and vendor redundancy, and you give up the managed sharing, Slack app and admin dashboard.

Match the model to the team job

Treat the table as a starting point.

Run your own team's top three jobs through two models before committing.

What no benchmark tells you about Grok Bot teams

The parts of a shared Bot that decide whether a team trusts it aren't on any leaderboard.

• Privacy: each person's chat is private, even from the owner and Managers.
• Usage: each reply uses the usage of the person it's answering.
• Approvals: in Slack, only the person who asked can answer a Review an action card.
• Safety gap: the teams docs say shared Bots in chats nobody can approve in skip Auto-review unless Enterprise enforcement is on.
• Plans: under The X Subscription, Premium, Super and Ultra include Grok Bot; Plus doesn't.

Check those before you worry about a tenth of a point.

There's also a timing point.

The feature changed in eight releases between v0.61.0 and v0.78.1, from Managers to team Vault logins.

A shared Bot's behaviour can move week to week even though nobody on your team touched a model setting.

That's another reason to judge it on your own jobs, not on a static score.

How to read GoldieBench numbers honestly

Every model gets the same fixed prompt, with no help, no iteration and no best-of-N picking.

Medals go to the top three scores per task, and the rank is the average across scored tasks.

Reference benchmarks like GDPval or Terminal-Bench on model pages are outside numbers with their source shown, not GoldieBench scores.

And provisional rows with only a handful of tasks, like LongCat-2.0 on 4 tasks, shouldn't be compared directly with 47-task rows.

The bottom line

Use Grok Bot's Shared Bot when the team needs one managed helper in Slack and the app.

Use GoldieBench to pick models for anything you run yourself, with Hermes MoA and Fusion at the top of the open options.

For the Agent OS that runs those models, join the AI Profit Boardroom.

Also On Our Network

• Agent Operatorsthe six-step setup order with every menu name and admin switch
• AI Income Deskwhether a shop, firm or small agency should give staff one shared Bot
• AI Tool Verdicta ten-criterion scorecard and a straight 7.6/10 verdict
• agentos.guidefive Agent OS prompts to brief, audit and roll out a shared Bot
• aiprofitboardroom.comthe complete Grok Bot Teams guide on the AI Profit Boardroom blog

Set up your Shared Bot for the team, benchmark your own top three jobs, and book a free strategy session if you want help choosing models around Grok Bot teams.