Grok Bot Phone Calls + Which AI Models To Pair With Them, With Real GoldieBench Scores

Grok Bot phone calls run on whichever model Grok Bot chooses, so the GoldieBench scores worth reading are for the models you use around each call: drafting the brief, preparing answers and summarising the transcript afterwards.

The Grok Bot FAQ is plain about it: Grok Bot picks the model for each task, there is no model picker, and it doesn't show the underlying model for a request.

So this page won't guess which model speaks on your calls.

Every score below comes from a live GoldieBench page I read on 11 November 2026, linked next to the number.

If you'd like my team to help plan your setup, book a free SEO strategy session.

What a Bot phone call involves, before any model talk

Per xAI's phone call docs, a Bot can phone a business on your behalf (the docs' examples are a table booking and an order check) and then tell you how it went.

Calls are US only and start from the Grok Bot desktop or phone app.

The first time, you accept calling terms that include only ringing US businesses and people who've consented to hear from you.

Before each call you see who the Bot will ring and choose Call or Cancel.

During the call there's a live transcript, End call, Take over, and timed question cards for anything the Bot can't answer.

Afterwards you get the transcript and the recording, with export and download options on desktop.

Why the model question has no answer here

xAI says Grok Bot routes each task to the backend model most likely to do it well, which can be its own model or a third-party one.

Your usage views record that as Grok Bot usage, not by provider or model.

That leaves the planning around a call as the place where your own model choice counts.

What GoldieBench measures, and what it leaves out

The methodology page sets the rules: the same fixed prompt for every model, one shot, and one self-contained HTML file as the output.

The Grok deep dive lists what the bench does not measure: multi-turn agent work, long-context recall and API latency.

A live phone conversation is multi-turn work with tight timing, which is exactly the part the bench doesn't touch.

So read these scores as a signal of how cleanly a model turns one clear instruction into a finished piece of work.

That matches writing a call brief or a summary template, not holding the conversation.

The scores worth knowing for call prep

The Grok row says #8 of 26 on its own benchmark page, while the GoldieBench leaderboard lists it seventh, with 43 of 47 tasks scored; I quote the 8.09 and rank from the deep dive.

The Pages category covers only three tasks on each of these pages, so treat it as a small sample.

The Grok row describes xAI's model as benched, and nothing in the Grok Bot docs ties it to what runs on your calls.

Drafting call briefs: where the Pages row matters most

A good brief is one instruction turned into a tidy, complete document, which is close to what the Pages tasks test.

Claude Opus 5, GPT-5.6 Sol and Fusion each post 8.5 in Pages against a field average of 7.7.

Grok posts 8.3 there, and Hermes MoA posts 8.0.

Any of the five clears the field average, so pick the one you already pay for and keep the brief short.

The structure matters more than the model: who to call, why, what to share, what not to agree to, and how to report back.

Reading transcripts afterwards: why the bench says little

Summarising a call transcript is a reading job, and the bench doesn't score reading or recall.

The Opus 5 and GPT-5.6 Sol pages list context windows of 1,000,000 and 1,050,000 tokens, and the Grok page lists 256,000.

Any of those comfortably holds a single call's transcript, so context isn't the deciding factor for one call.

If you batch a week of exported transcripts into one review, the bigger windows give you more room.

Panels versus single models for prep work

Fusion tops the board at 8.59 with 21 golds, and Hermes MoA sits third at 8.17; both are multi-model panels.

Claude Opus 5 at 8.27 is the best single model in this table.

For a two-paragraph call brief, a panel is more machinery than the job needs.

For a reusable library of briefs covering every supplier you deal with, a panel's extra pass can be worth it.

A pairing plan for two small businesses

A gym owner checking an equipment repair

1. Ask your chosen model to draft a brief: which repair firm, the machine, the fault and the visit window you'll accept.
2. Paste the brief into the chat with your Bot and confirm the firm on the Call card.
3. Answer any question card about access or parking yourself.
4. Export the transcript on desktop and have the same model turn it into a one-line job note.

A print shop chasing a paper delivery

1. Have a model prepare the order details and the questions the supplier is likely to ask.
2. Ask your Bot to make the call and press Call when the card shows the supplier.
3. If the supplier offers a different stock, press Take over and decide on the line.
4. Feed the transcript to your model for a dated summary in the order file.

For more on using a Bot across a business, see our Grok Bot for business pairing guide.

Gaps no score can fill

No benchmark tells you what the call feature itself leaves undocumented.

• Call length or minute limits, and how many calls you can place.
• The caller ID the business sees, and whether it hears that an AI is calling.
• Supported languages and how long recordings are kept.

The Grok Bot changelog adds fixes rather than limits: v0.79.2 on 5 November 2026 brought an editable "Send this to your call?" card, and v0.79.3 on 6 November added No answer for busy lines and Copy conversation ID on failed calls.

It has no launch entry for calls, so there's no launch date to quote.

Bottom line

Let Grok Bot handle the call, and spend your model choice on the brief and the summary.

For that work, any model in the table above clears the Pages field average, with Claude Opus 5, GPT-5.6 Sol and Fusion level at the top of that category.

Keep calls to US businesses and people who've consented, and press Call only after checking the card.

To see members' prompt libraries for this kind of prep, the AI Profit Boardroom has them, and you can join the AI Profit Boardroom here.

Also On Our Network

• Agent Operatorsthe six-step setup from Vault entry to an exported call transcript
• AI Income Deskfour small-business errands a Bot can phone through, and the rules
• AI Tool Verdicta scored verdict on control, records, reach and fit, rated 7 out of 10
• agentos.guidefour copy-paste prompts that brief, prep and log every Bot call

Pick a model that clears the Pages average, keep the brief tight and the approval habit tighter, and if you want help building the rest of the system, book a free SEO strategy session to plan the models around your Grok Bot phone calls.