⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 7 min read

Grok Bot Outlook And GoldieBench: Which AI Models To Pair For The Files Your Bot Attaches, With Live Scores

Grok Bot chooses its own model, so the scores matter for the templates and files you build around it. Here is the live Pages tab and how to use it.

Illustrated leaderboard of glowing bars beside an envelope with a paperclip, linked by a gold approval tick

Grok Bot Outlook work runs on whatever model Grok Bot chooses for each task, so the benchmark question that matters is which models you pair around the Bot for the files it attaches, not which model reads your mail.

This page uses the live GoldieBench leaderboard as I read it on 14 November 2026, and xAI's own docs for the Outlook facts.

Those facts in brief: the Bot can read emailed login codes from Outlook, bring attachments onto its computer, and attach files to Outlook drafts, per the computer and apps page and the files page.

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

Grok Bot picks its own model

The Grok Bot FAQ is direct about this: Grok Bot picks the model for each task and offers no model picker.

It also says the underlying model isn't shown for a request.

So no GoldieBench row is a Grok Bot Outlook score, and that includes the row named Grok.

Where the bench helps is the work you do beside the Bot: building quote pages, report pages and templates that later end up as attachments on Outlook drafts.

What GoldieBench measures, and what it doesn't

The methodology page gives every model the same fixed prompt, one shot, with no help and no best-of-several picking.

Each result must be a single self-contained HTML file that runs in a browser.

Scores are Julian's own 0 to 10 numbers, and the leaderboard rank is a model's average across the tasks it was scored on.

The categories are Games (23 tasks), Visuals (9), Sims (12) and Pages (3).

Nothing on the bench grades email writing, reading attachments or handling an inbox, so read these numbers as build quality only.

The overall leaderboard, read 14 November 2026

RankModelAvg / 10Tasks scored
1Fusion8.5947 of 47
2Claude Opus 58.2750 of 50
3Hermes MoA8.1747 of 47
4GPT-5.6 Sol8.1650 of 50
5Claude Fable 58.1047 of 47
6Qwen 3.88.1045 of 45
7Grok8.0943 of 47
8MiniMax M37.9747 of 47

Every number is from the GoldieBench leaderboard, which listed 50 tasks, 33 models and 1,310 live demos that day.

The top seven sit inside half a point, so treat small gaps as noise.

LongCat-2.0 shows 8.12 but is marked provisional on four tasks, so it carries no rank.

The Pages tab is the closest fit for Outlook attachments

An Outlook reply from a Bot often carries a file you designed once and reuse: a quote page, a one-page report, an order summary.

The Pages category tests single-file web pages, and its three tasks are named Landing, Aipbpromo and Webos on the matrix.

ModelPages avg / 10Overall avg / 10
Fusion8.508.59
Claude Opus 58.478.27
GPT-5.6 Sol8.478.16
Qwen 3.88.408.10
Grok8.338.09
MiMo-V2.6 Pro8.206.35
MiniMax M38.177.97
Muse Spark 1.28.177.55
Gemini 3.6 Flash8.137.08

All Pages figures come from the Pages tab of the same leaderboard, each over three scored tasks.

Three tasks is a small sample, so use this tab to shortlist, then build your own template and judge it.

Two rows stand out: MiMo-V2.6 Pro and Gemini 3.6 Flash do far better on Pages than their overall average suggests.

Claude Fable 5, by contrast, sits fifth overall but scores 7.57 on Pages.

Where the three Grok rows sit

The bench lists three xAI rows, and none of them is Grok Bot.

  • Grok: 8.09 overall, 43 of 47 tasks scored, marked best at Games, and 8.33 on Pages.
  • Grok 4.7: 7.15 overall across 20 scored tasks, with no Pages score.
  • Grok 4.6: 5.95 overall across 20 scored tasks, also with no Pages score.

It's tempting to read the top Grok row as a preview of the Bot, but xAI doesn't say which model Grok Bot runs for a given request.

Treat the Grok rows like any other model: useful for building templates, silent about how the Bot handles your mail.

How to pair a model with Grok Bot Outlook

  1. Pick one model from the top of the Pages tab and build your template once, as a single file.
  2. Save it where your Bot can reach it, then connect Outlook under Connect apps in Grok Bot.
  3. Ask the Bot to fill the template from an email attachment it brings onto its computer.
  4. Have it attach the result to an Outlook draft and stop.
  5. Check the draft card, which lists the attached files since v0.79.2 (5 November 2026, per the changelog), then press Send.

The model you chose shaped the template; Grok Bot handles the mail part with its own model choice.

Keep sending under your control

The approvals page says Auto Review checks each action, including a connected-app call, before it runs.

Add a Custom Rule such as "send an email from my Outlook" set to Ask first, and every send will wait for you.

Auto Review only counts your own messages as requests, so text inside a customer's email can't approve a send.

No benchmark score changes that; it's a setting, not a model property.

Work accounts need IT approval first

A work or school Microsoft 365 account can block Outlook from connecting until the organisation's admin approves Grok Bot.

From a computer, the Microsoft page shows an Approval link for your IT admin; the phone app shows the message without the link.

The admin's step is tenant-wide admin consent in Microsoft Entra ID.

On a team, a blocked Outlook shows Disabled by team admin, set by the team's connector policy.

If your team runs a shared Bot in Microsoft chat as well, see the Grok Bot Microsoft Teams model guide.

What the numbers can't tell you

  • How well a model writes a polite supplier reply.
  • How accurately it reads a messy spreadsheet attachment.
  • How Grok Bot itself performs, since its model isn't shown.
  • Whether a template prints cleanly as a PDF.

For those, run a test email to yourself and judge the draft card.

A two-template test before you commit

Scores shortlist; your own test decides.

  1. Pick the top two models from the Pages tab.
  2. Give both the same brief for your quote or report template.
  3. Have Grok Bot fill each one from the same email attachment and attach both to one draft addressed to you.
  4. Open the email, compare the two files side by side, and keep the better template.

This mirrors the bench's own rule of one fixed prompt for every model, applied to your real work.

My verdict on models for Grok Bot Outlook

Use the Pages tab, not the overall rank, when you choose a model to build the files your Bot attaches.

Fusion, Claude Opus 5 and GPT-5.6 Sol lead that tab on the 14 November read, with Qwen 3.8 and Grok close behind.

Leave the mail work to Grok Bot, guard sending with an Ask first rule, and judge every result on the draft card.

For the trainings and the Agent OS behind my own setup, join the AI Profit Boardroom.

Also On Our Network

Shortlist from the Pages tab, talk your stack over on a free SEO strategy session, and learn the build inside the AI Profit Boardroom so every template makes Grok Bot Outlook drafts worth sending.

FAQ

Which model does Grok Bot use for Outlook tasks?

The Grok Bot FAQ says it picks the model for each task, with no picker, and doesn't show the underlying model. No GoldieBench row is a Grok Bot score.

Which GoldieBench category fits Outlook work?

Pages, because it tests single-file web pages, the closest match to the quote or report templates a Bot attaches to drafts.

Who leads the Pages tab?

On the 14 November 2026 read: Fusion 8.50, Claude Opus 5 and GPT-5.6 Sol 8.47 each, Qwen 3.8 8.40 and Grok 8.33.

Does GoldieBench test email writing?

No. Every task is a one-shot single HTML file, so the scores say nothing about email tone or reading attachments.

Is the Grok row on the leaderboard the same as Grok Bot?

No. It is a model row scored on GoldieBench tasks. Grok Bot chooses its own model per task and doesn't say which.

How do I check a Bot-built attachment before it goes out?

Ask for an Outlook draft and stop. The draft card lists the attached files, and nothing sends until you press Send.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week