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

GoldieBench Blog · 7 min read

Grok Bot Meeting: Which AI Models To Run On Your Meeting Notes, With Real GoldieBench Scores For Builders And Checkers

Grok Bot picks its own model in the call, so the scores that matter belong to the builder and checker models you run on the transcript afterwards.

A dark navy leaderboard of glowing bars beside an abstract video call grid, with gold lines carrying note cards from the call into two highlighted bars

A grok bot meeting gives you notes, but not a model choice, so the scores worth reading are for the model you run afterwards to turn those notes into recap pages, client briefs and follow-up checks.

The Grok Bot FAQ says there's no model picker: Grok Bot chooses the backend model for each task and doesn't show which one it used.

Every GoldieBench number below was read from live goldiebench.com pages on 12 November 2026, and each one names the page it came from.

For a plan built around your own calls, book a free SEO strategy session with my team.

What the Bot does in a meeting, briefly

Per xAI's meetings documentation, a Bot joins Google Meet, Zoom or Microsoft Teams once the host admits it, listens, takes notes, answers in the meeting chat when addressed by name, and sends a summary and action items when the call ends.

Meetings is available on Team, Enterprise and Ultra plans, and Enterprise admins can switch it off for everyone (team settings docs).

You can ask the Bot later for the summary or transcript of any past meeting.

That transcript is the raw material your own model works on.

What a GoldieBench score measures, and what it doesn't

The methodology page says each task is one fixed prompt, one shot, with a single self-contained HTML file out.

Scores are Julian's own 0 to 10 numbers from his head-to-head guides, and medals go to the top three per task.

So a score tells you how well a model ships a working page or app from one prompt.

It doesn't measure transcription, summarising or note accuracy, and nothing on the bench tests Grok Bot itself.

Use the numbers to pick the model that builds things from your meeting notes, not to judge the notes.

The overall leaders on the home leaderboard

RankModelAvg / 10Tasks scoredPage read
#1Fusion8.5947/47Home leaderboard
#2Claude Opus 58.2750/50Home leaderboard
#3Hermes MoA8.1747/47Home leaderboard
#4GPT-5.6 Sol8.1650/50Home leaderboard
#5Claude Fable 58.1047/47Home leaderboard

The home page ranks by the average 0 to 10 score across every task a model was scored on.

For a meeting workflow, the overall top two are the safe default when you want one model for everything after the call.

The models tagged best at Pages

Pages is the closest bench category to meeting follow-up work, because a recap page or a client tracker is a page.

The leaderboard lists Pages as 3 tasks, and tags these models "best at Pages":

RankModelAvg / 10Medals
#4GPT-5.6 Sol8.162 gold, 9 silver, 10 bronze
#11GLM-5.27.775 gold
#18Gemini 3.6 Flash7.081 gold, 2 silver, 1 bronze
#19Claude Sonnet 57.014 bronze

GPT-5.6 Sol is the pick if you want the strongest all-rounder that also leans towards pages.

GLM-5.2, Gemini 3.6 Flash and Claude Sonnet 5 also carry the tag, which makes any of them a fair test as your checker model.

What the Grok page shows, for context

The Grok model page lists an 8.09/10 average across 47 tasks, with 5 gold, 1 silver and 1 bronze medal and a 256,000-token context.

On its page-type tasks it shows Landing Page at 9.0 and Webos Page at 9.0, both gold, and Aipbpromo Page at 7.0.

The home leaderboard places Grok at #7, with 43 of 47 tasks scored.

These are scores for the Grok model on the bench, not a statement about what Grok Bot runs; the FAQ says the Bot's model can change from task to task.

Context length matters for long transcripts

A long client call produces a long transcript, and your checker model has to hold all of it at once.

The home leaderboard lists MiniMax M3 at #8 with a 7.97/10 average and a 1,048,576-token context window in its reference column.

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

If your calls run long or you want to check several meetings in one pass, put context length next to the score when you choose the checker.

For a short weekly stand-up, any model in the tables above has room to spare.

A three-step model setup for meeting follow-up

  1. Bot: joins the call, takes notes and hands you the summary, action items and transcript.
  2. Builder model: one of the overall top two, or GPT-5.6 Sol, turns the checked notes into a recap page or tracker.
  3. Checker model: a different model compares the recap with the transcript and lists anything unsupported.

The Agent OS inside the AI Profit Boardroom is where Julian runs his own builds, per the methodology page, so it's a natural home for steps two and three.

Using two different models for build and check means one model's blind spot is less likely to pass through both steps.

If you have Pages, the Bot already writes a longer report on a page; Pages is rolling out gradually and isn't on Enterprise teams yet, which is where your own builder model fills the gap.

Prompts for the builder and the checker

Builder:
Here is the summary, action items and transcript from today's client call.
Build one self-contained HTML recap page: a table of actions with owner and date,
a short list of decisions, and open questions. No external data.
Checker:
Compare this recap page with the transcript.
List every date, quantity or commitment on the page the transcript doesn't support.

The builder prompt matches the bench's own format, one prompt to one HTML file, so the scores above apply directly.

Small-business examples

An agency after a client kickoff

  1. Disable Interactivity on the scheduled meeting so the Bot only listens.
  2. Feed the transcript to your builder model for a client-facing recap page.
  3. Run the checker before you send the link.

An accountant after a year-end review

  1. Ask the Bot for the action items.
  2. Have the builder make a document-request checklist page.
  3. Have the checker confirm each document against the transcript.

A plumbing firm after a property manager call

  1. Ask the builder for a job tracker grouped by address.
  2. Run the checker on dates before the tracker goes to the client.

Meeting rules no model choice changes

  • The host admits the Bot every time.
  • Tell attendees a Bot is transcribing, and follow your own consent and recording rules.
  • Your credentials stay out of the call, and any request the Bot declines comes to you with the name of whoever asked.
  • Meetings use the settings of the person who started them.

Per the security docs, a Bot owns no credentials, and the model never sees your logins.

For Bots inside Microsoft Teams chats rather than calls, see our Grok Bot Microsoft Teams breakdown.

Where to recheck the numbers

Scores move when new models land, so recheck the models page and the home leaderboard before you lock in a builder.

Meeting features change on the Grok Bot changelog.

When either page changes, rerun the builder and checker prompts on one old transcript and compare the output before you switch models.

For the Agent OS and the trainings that run this workflow end to end, join the AI Profit Boardroom.

Also On Our Network

Pick one builder and one checker from the tables and run both on your next transcript, and book a free SEO strategy session if you want my team to plan the workflow around every grok bot meeting.

FAQ

Can I choose which model Grok Bot uses in a meeting?

No. The FAQ says there is no model picker and Grok Bot chooses the backend model per task. Choose the models you run on its notes instead.

Does GoldieBench test meeting transcription?

No. The methodology page describes one-prompt, one-shot HTML builds scored 0 to 10. It measures building, not transcribing.

Which model tops the home leaderboard?

Fusion, at #1 with 8.59/10 across 47 scored tasks, with Claude Opus 5 second at 8.27 across 50.

Which models are tagged best at Pages?

On the home leaderboard, GPT-5.6 Sol, GLM-5.2, Gemini 3.6 Flash and Claude Sonnet 5 carry the "best at Pages" tag, among others.

Why use a separate checker model?

A different model reviewing the recap against the transcript lowers the chance that one model's mistake passes both steps.

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