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

GoldieBench Blog · 6 min read

Grok Bot Vault And Which AI Models To Pair With It, Using Real GoldieBench Scores For The Pages Around It

Grok Bot chooses its own model, so the leaderboard matters for the pages and agents you build next to Vault.

Illustrated leaderboard of glowing gold bars beside a locked vault door connected by thin light lines

Grok Bot Vault has no model setting, so the GoldieBench scores worth reading are the ones for the models that build the tools around it: access maps, onboarding pages and admin dashboards.

The Grok Bot FAQ says Grok Bot picks the model for each task and offers no model picker.

So this page doesn't rank models for Vault itself.

It ranks them for the jobs you do next to Vault, using scores I read on live GoldieBench pages on 13 November 2026.

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

Vault in one minute, for context

The Vault docs describe three saved item types: logins, cards and contact info.

Bots use them to sign in and fill checkout forms while the passwords, card numbers and security codes stay hidden from the Bot.

API keys go in a masked secure secret card, and shared Bots keep their own Secrets.

Everything is write-only once saved.

None of that depends on which model answers a request.

Why model choice moves outside the Bot

The FAQ says the backend can be a first-party or a third-party model, and it can change from one request to the next.

Grok Bot doesn't display which model handled a request, and you can't route a request to or away from one.

That's fine for credential handling, because the protection lives in Vault and Secure Forms, not in the model.

Where you do choose models is in your wider stack: the agents in your Agent OS that write the docs, pages and trackers your team uses alongside the Bot.

How GoldieBench scores work

The methodology page sets the rules.

  • Every model gets the same fixed prompt, one shot, with no iteration and no best-of-N picking.
  • Each result must be a single self-contained HTML file that runs in a browser.
  • Scores are Julian's own 0 to 10 judgements, and medals go to the top three per task.

No task on the bench is a password manager or a sign-in flow.

The nearest category is Pages, which covers the kind of internal pages a Vault rollout produces.

The leaderboard top five

From the GoldieBench leaderboard, read on 13 November 2026:

RankModelAverageBest at
#1Fusion8.59Games
#2Claude Opus 58.27Sims
#3Hermes MoA8.17Visuals
#4GPT-5.6 Sol8.16Pages
#5Claude Fable 58.10Games

The leaderboard lists 33 models and 50 tasks.

For page-style work, GPT-5.6 Sol is the one marked best at Pages, with 50 of 50 tasks scored.

Grok's own numbers on the bench

Grok sits at #7 on the leaderboard with an 8.09 average and 43 of 47 tasks scored.

The Grok deep-dive page, data refreshed 27 October 2026, lists its three Page scores:

TaskGrok scorePage read
Landing Page9.0, goldgoldiebench.com/benchmarks/grok
Webos (Web-OS desktop)9.0, goldgoldiebench.com/benchmarks/grok
Aipbpromo7.0goldiebench.com/benchmarks/grok

Two golds in three page tasks is a strong page record.

That said, the Grok on the bench is a model you call yourself, and Grok Bot doesn't tell you which model it uses for a task.

Jobs around Vault, and the score that matters for each

  • Access map page: a single page listing which agent uses which login, by website and hint, with no secrets. Read Page scores.
  • Staff onboarding guide: how to use Secure Forms, Take over and I'm done. Read Page scores again.
  • Rotation tracker: a checklist of keys to Replace each month. A simple page task.
  • Everything else in your agent team: the overall average is the better guide.

Never paste real secrets into any of these builds.

The Vault docs say passwords, one-time codes and API keys never belong in ordinary chat, a Shell command or a file.

My pairing setup, by job

  • Sign-in and checkout: Grok Bot itself, with Vault, Secure Forms and your Ask first rules from the approvals page.
  • Access map and onboarding pages: GPT-5.6 Sol, the leaderboard's best at Pages, or Grok, with its 9.0 Landing Page score.
  • The wider agent team: Fusion, Claude Opus 5 or Hermes MoA, the top three overall.

Start with one page model and one build.

Check the access map against your real Settings, Vault list, then add the next model.

A sample access map, column by column

Here's the page I'd brief a model to build first.

It lists credentials by name only, so it's safe to share with the team.

ColumnWhat goes in itWhere the fact comes from
WebsiteThe site a login is forYour Settings, Vault list
HintWhich agent or job uses itThe entry's Hints field
StorePersonal Vault, Team Bot login or secure cardWhere you saved it
OwnerWho can replace the valueYou, or the Team Bot managers
Hand-over stepsSSO, passkey or QR code the Bot passes to youYour own notes after each Take over

Give the model the column list and a plain description of each agent's job.

Leave every password, key and card number out of the prompt.

Reading the medals for page work

Medals on GoldieBench are per task: gold for the top score on that task, then silver and bronze.

A model can be best at Pages overall without winning every page task, and a model ranked lower overall can still take gold on a single page build.

Grok shows both sides of that: #7 overall, yet gold on Landing Page and Webos.

For an access map or onboarding guide, a page-task gold is the more useful signal than the overall rank.

Click into the task on the site and play each model's build before you choose, since every demo on the bench is live.

Team Bots change the picture

If your business runs shared agents, remember that Team Bots don't use your personal Vault.

Each one keeps its own logins, visible to managers only, and its own Secrets, where the Bot sees the description but never the value.

Your access map page should show personal and team credentials in separate sections for that reason.

The Grok Bot shopping model guide on this site covers the buying side.

What these scores don't tell you

GoldieBench measures one-shot builds, not security.

A high Page score says a model can ship a clean internal page from one prompt.

It says nothing about how Vault stores data, and the Vault page itself doesn't name an encryption method either.

Use the bench for builds, and the xAI docs for how credentials are handled.

Members of the AI Profit Boardroom get the prompts I use to run these pairings.

Also On Our Network

Pick one page model, build your access map, and talk to my team or join the AI Profit Boardroom to pair the right models with Grok Bot Vault.

FAQ

Which model does Grok Bot Vault use?

None you can choose. The Grok Bot FAQ says Grok Bot picks the model per task, with no model picker.

Which GoldieBench model is best at pages?

On the leaderboard read on 13 November 2026, GPT-5.6 Sol is marked best at Pages, with an 8.16 average.

How did Grok score on page tasks?

The Grok deep-dive page lists 9.0 gold for Landing Page, 9.0 gold for Webos and 7.0 for Aipbpromo.

Does GoldieBench test password security?

No. It scores one-shot HTML builds. For how credentials are handled, read the xAI Vault docs.

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